Embed Size (px)
DESCRIPTION
Research dissertation
Citation preview
THE APPLICATION OF IMAGE PROCESSING AND MACHINE
LEARNING TECHNIQUES FOR DETECTION AND CLASSIFICATION
OF CANCEROUS TISSUES IN DIGITAL MAMMOGRAMS
JAWAD NAGI
FACULTY OF COMPUTER SCIENCE AND
INFORMATION TECHNOLOGY
UNIVERSITY OF MALAYA
KUALA LUMPUR
2011
THE APPLICATION OF IMAGE PROCESSING AND MACHINE
LEARNING TECHNIQUES FOR DETECTION AND CLASSIFICATION
OF CANCEROUS TISSUES IN DIGITAL MAMMOGRAMS
BY
JAWAD NAGI
DISSERTATION SUBMITTED IN FULFILMENT
OF THE REQUIREMENTS
FOR THE DEGREE OF MASTER OF COMPUTER SCIENCE
FACULTY OF COMPUTER SCIENCE
AND INFORMATION TECHNOLOGY
UNIVERSITY OF MALAYA
KUALA LUMPUR
2011
ii
UNIVERSITY OF MALAYA
ORIGINAL LITERARY WORK DECLARATION
Name of Candidate: Jawad Nagi (I.C./Passport No: AD9990581)
Registration/Matric No: WGA080040
Name of Degree: Masters of Computer Science
Title of Project Paper/Research Report/Dissertation/Thesis (“this Work”): The Application of Image Processing and Machine Learning Techniques for
Detection and Classification of Cancerous Tissues in Digital Mammograms
Field of Study: Medical Informatics
I do solemnly and sincerely declare that: (1) I am the sole author/writer of this Work; (2) This Work is original; (3) Any use of any work in which copyright exists was done by way of fair dealing and
for permitted purposes and any excerpt or extract from, or reference to or reproduction of any copyright work has been disclosed expressly and sufficiently and the title of the Work and its authorship have been acknowledged in this Work;
(4) I do not have any actual knowledge nor do I ought reasonably to know that the making of this work constitutes an infringement of any copyright work;
(5) I hereby assign all and every rights in the copyright to this Work to the University of Malaya (“UM”), who henceforth shall be owner of the copyright in this Work and that any reproduction or use in any form or by any means whatsoever is prohibited without the written consent of UM having been first had and obtained;
(6) I am fully aware that if in the course of making this Work I have infringed any copyright whether intentionally or otherwise, I may be subject to legal action or any other action as may be determined by UM.
Candidate’s Signature Date:
Subscribed and solemnly declared before,
Witness’s Signature Date:
Name:
Designation:
iii
ABSTRACT
Breast cancer is one of the most common kinds of cancer, as well as the leading
cause of mortality among women. Mammography is currently the most effective
imaging modality for the detection of breast cancer and the diagnosis of the
anomalies which can identify cancerous cells. Retrospective studies show that, in
current breast cancer screenings approximately 15 to 30 percent of breast cancer
cases are missed by radiologists. With the advances in digital image processing
techniques, it is envisaged that radiologists will have opportunities to decrease this
margin of error and hence, improve their diagnosis.
Digital mammograms have become the most effective techniques for the detection
of breast cancer. The goal of this research is to increase the diagnostic accuracy of
image processing and machine learning techniques for optimum classification
between malignant and benign abnormalities in digital mammograms by reducing
the number of misclassified cancers. In this research, digital mammography images
are obtained from Malaysian patients who are treated at the University of Malaya
Medical Centre (UMMC) from 2008 to 2010. This database consists of standard
images of dense, fatty and fatty-glandular breasts, which are classified into three
categories: normal, benign and malignant, using the results obtained from biopsies.
Image processing techniques are applied in this research to enhance the
mammogram images for the computerized detection of breast cancer. Image
processing algorithms used for mammogram image processing include
morphological operations and thresholding techniques. As the pectoral muscle in
digital mammograms can bias the detection results, it should be suppressed from
iv
the mammograms. This research employs a seeded region growing technique for
the segmenting the breast tissue from the pectoral muscle.
Malignant and benign abnormalities are selected from the segmented images using
the Ground Truth (GT) data and markings obtained from the radiologists’
interpretation of the mammography datasets, which correspond to the Regions of
Interest (ROIs) or abnormal regions (samples). Texture based features are
extracted from the ROI samples using Gray Level Co-Occurrence Matrices (GLCMs).
For the purpose of pattern classification between malignant and benign samples,
the optimum subset of texture features are modeled using a Support Vector
Machine (SVM). The SVM is trained using two-thirds of the total samples where the
remaining one-third of samples are used for testing and validation. The binary
classification accuracy of the developed system is measured using the Receiver
Operating Characteristic (ROC) analysis with performance measures such as
sensitivity, specificity and the Area Under the Curve (AUC). To perform a
comparative study, machine learning algorithms other than the SVM, namely,
Artificial Neural Networks (ANNs) are evaluated in this research.
The experimental results obtained from the system developed in this research
prove to be beneficial for the automated detection of breast cancer. The proposed
technique will improve the diagnostic accuracy and consistency of the radiologists’
image interpretation in the diagnosis of breast cancer. The resulting computerized
breast cancer detection system will subsequently act as a second reader after the
manual detection by the radiologist and it is believed that this would aid the
radiologist in the mammogram screening process.
v
ACKNOWLEDGEMENTS
First and foremost, I wish to thank God for giving me strength and courage to
complete this thesis and research, and also to those who have assisted and
inspired me throughout this research.
There are so many people to whom I am indebted for their assistance during my
endeavors to complete my Masters candidature in Computer Science at University
of Malaya (UM). First and foremost, I would like to express my gratitude to my
supervisor Assoc. Prof. Datin Dr. Sameem Abdul Kareem from University of
Malaya, Malaysia, whose invaluable guidance and support was very helpful
throughout my research. A similar level of gratitude is due to Dr. Farrukh Hafiz
Nagi and Mr. Syed Khaleel Ahmed from Universiti Tenaga Nasional, Malaysia. It is
unlikely that I would have reached completion without their encouragement and
support.
I express my appreciation to everyone involved directly and indirectly to the
success of this research. Last but not least, my family for their understanding,
support, patience, and encouragement. Thank you for all the support, comments
and guidance.
vi
DEDICATION
This thesis is dedicated to my father, who taught me that the best kind of
knowledge to have is that which is learned for its own sake. It is also dedicated to
my mother, who taught me that even the largest task can be accomplished if it is
done one step at a time.
vii
TABLE OF CONTENTS
DECLARATION
Page
ii
ABSTRACT iii
ACKNOWLEDGEMENTS v
DEDICATION vi
TABLE OF CONTENTS vii
LIST OF FIGURES xiii
LIST OF TABLES xix
LIST OF ABBREVIATIONS xx
CHAPTER 1 – INTRODUCTION
1.0 Overview
1.1 Problems and Motivation
1.2 Research Objectives and Scope
1.3 Research Significance and Contribution
1.4 Research Methodology and Proposed Approach
1.5 Benefits of Image Processing and Machine Learning Techniques
1.6 Thesis Overview
1
6
13
14
16
17
19
CHAPTER 2 – DIGITAL MAMMOGRAPHY
2.0 Overview
2.1 Breast Anatomy and Cancer
2.1.1 Calcifications
2.1.2 Mass Lesions
2.2 Breast Tumors
2.2.1 Non-Cancerous Breast Tumors
2.2.2 Cancerous Breast Tumors
2.2.2.1 Non-Invasive Breast Cancer
2.2.2.2 Invasive Breast Cancer
2.3 Differentiating Between Breast Tumors
2.3.1 Mammography
2.4 Screening for Breast Cancer
22
22
25
25
28
29
29
30
31
33
35
37
viii
2.4.1 Errors In Screening
2.5 Imaging Modalities
2.5.1 Digital Mammography
2.5.2 Ultrasonography
2.5.3 Magnetic Resonance Imaging
2.6 Mammogram Analysis Using Digital Mammography
2.6.1 Breast Positioning in Digital Mammography
2.6.2 Breast Regions in Digital Mammograms
2.6.3 Types of Breast Tissues
2.6.4 Present Clinical Protocol
2.6.4.1 BI-RADS Descriptors and Assessment
2.6.4.1.1 BI-RADS Mass Descriptors
2.6.4.1.2 BI-RADS Assessment Categories
2.6.5 Mammogram Interpretation
2.7 Summary
43
44
45
46
46
46
47
48
51
52
53
53
54
54
55
CHAPTER 3 – ELEMENTS OF COMPUTER-AIDED DETECTION
3.0 Overview
3.1 Computer-Aided Detection Systems
3.2 Review of Computerized Breast Cancer Detection Techniques
3.2.1 Detection of Microcalcifications/MCCs
3.2.2 Detection of Mass Lesions
3.2.2.1 Central Mass
3.2.2.2 Spicules
3.2.2.3 Normal and Abnormal Regions
3.3 Computerized Detection of Breast Cancer
3.3.1 Image Preprocessing
3.3.2 Image Segmentation
3.3.3 Feature Extraction
3.3.4 Feature Selection
3.3.5 Classification
3.3.6 Performance Evaluation
3.4. Fundamentals of Digital Image Processing
3.4.1 Representation of a Digital Image
57
57
58
61
64
65
67
68
69
70
72
73
76
78
80
85
85
ix
3.4.1.1 Range of Intensity Values
3.4.2 Histogram
3.4.2.1 Uses of Histogram
3.4.2.2 Histogram Normalization
3.4.2.3 Histogram Equalization
3.4.3 Image Segmentation
3.4.3.1 Thresholding Techniques
3.4.3.2 Boundary-based Methods
3.4.3.3 Region-based Methods
3.4.3.3.1 Selected Segmentation Technique
3.4.3.3.2 Seeded Region Growing
3.4.4 Morphological Operations
3.4.4.1 Dilation
3.4.4.2 Erosion
3.4.4.3 Morphological Opening and Closing
3.5. Textural Extraction and Analysis
3.5.1 Introduction to Texture Analysis
3.5.2 Texture Analysis Applied to Digital Mammography
3.5.2.1 Gray-level Co-occurrence Matrices (GLCMs)
3.5.2.2 Law’s Texture Filter
3.5.2.3 Local Binary Patterns (LBPs)
3.5.3 Comparison of Texture Analysis Techniques
3.5.3.1 Selected Texture Extraction Technique
3.5.3.1.1 Introduction to GLCMs
3.5.3.1.2 GLCM Texture Descriptors
3.6. Summary
87
88
89
90
90
91
92
94
96
97
98
102
103
104
104
105
105
107
109
111
113
113
115
116
119
120
CHAPTER 4 – PATTERN RECOGNITION AND FEATURE SELECTION
4.0 Overview
4.1 Machine Learning
4.1.1 The Act of Learning
4.1.2 Learning Pattern Classification
4.1.2.1 Learning from Data
4.1.2.2 Supervised Learning
122
122
123
126
126
128
x
4.1.2.3 Pattern Classification Issues
4.1.3 Validation Techniques
4.1.3.1 Hold-out Method
4.1.3.2 Cross-validation
4.1.3.3 Leave-one-out Method
4.2 Support Vector Machine (SVM)
4.2.1 Statistical Learning Theory
4.2.2 Linking Statistical Theory to SVM
4.2.3 Linear SVM
4.2.3.1 Separable Case
4.2.3.2 Non-separable Case
4.2.4 Non-linear SVM
4.2.5 Implementation of SVM
4.2.5.1 Sequential Minimal Optimization
4.3 Artificial Neural Networks (ANNs)
4.3.1 Back-Propagation Neural Network (BPNN)
4.3.2 Online-Sequential Extreme Learning Machine (OS-ELM)
4.4 Recursive Feature Elimination (RFE)
4.4.1 Feature Ranking Using F-score
4.4.2 SVM-RFE Using Random Forest (RF)
4.5 Summary
130
132
133
134
136
137
138
142
144
144
147
149
152
153
155
155
158
162
163
166
168
CHAPTER 5 – FRAMEWORK MODELING
5.0 Overview
5.1 Proposed Framework
5.2 Research Methodology and Implementation
5.2.1 Data Acquisition
5.3 Mammogram Image Processing
5.3.1 Image Preprocessing
5.3.1.1 Noise Removal
5.3.1.2 Radiopaque Artifact Suppression
5.3.2 Image Segmentation
5.4 Texture Feature Extraction and Selection
5.4.1 Region of Interest (ROI) Selection
169
169
172
174
179
179
179
182
189
208
208
xi
5.4.1.1 Necessity for Mammogram Image Processing
5.4.2 Texture Feature Extraction
5.4.3 Texture Feature Selection
5.4.4 Feature Normalization
5.5 Classification Engine Development
5.5.1 Feature Labeling and Adjustment
5.5.2 Training and Testing Data Separation
5.5.3 SVM Model Development
5.5.3.1 SVM Parameter Optimization
5.5.3.2 Probability Estimation
5.5.3.3 SVM Training
5.5.3.4 SVM Testing and Validation
5.5.3.5 Logic System for False Positive (FP) Reduction
5.6 Summary
211
213
218
220
221
221
222
222
223
226
227
229
230
233
CHAPTER 6 – EXPERIMENTAL RESULTS AND DISCUSSION
6.0 Overview
6.1 Experimental Results of Proposed Framework
6.1.1 Image Segmentation Performance Indices
6.1.1.1 Image Segmentation Results
6.1.2 Feature Selection Results
6.1.2.1 Discussion of F-score Results
6.1.3 SVM Training Validation
6.1.3.1 Discussion of SVM Training Results
6.1.4 SVM Testing and Validation
6.1.4.1 SVM Classification Results
6.1.4.1.1 Optimum ROI Size Selection
6.1.4.1.2 SVM Testing
6.1.4.1.3 False Positive Reduction Results
6.1.4.2 Discussion of SVM Classification Results
6.2 Comparison of Proposed Framework with Other Techniques
6.2.1 Experimental Results of Compared Techniques
6.2.2 Discussion of Compared Models
6.3 Summary
234
234
234
237
239
240
241
243
244
245
245
247
250
251
253
253
258
264
xii
CHAPTER 7 – CONCLUSION AND FUTURE WORK
7.0 Overview
7.1 Benefits of the Developed System
7.2 Contribution and Significance of Research
7.3 Achievement of Research Objectives
7.4 Impact and Significance to Radiologists
7.5 Future Expansion and Recommendations
7.5.1 SVM Parameter Tuning using Genetic Algorithm (GA)
7.5.2 Implementation of Multi-Scale RBF Kernel
7.5.3 Evaluating Other Texture Approaches
7.6 Conclusion
266
266
268
271
273
275
276
277
278
278
REFERENCES 280
APPENDICES 316
Appendix A: Data Modeling and Analysis 317
Appendix B: SVM Training and Testing 323
Appendix A: LIBSVM Copyright Notice 328
Appendix A: List of Publications 329
BIODATA OF AUTHOR 332
xiii
LIST OF FIGURES
Figure No. Page
1.1 Incident and mortality statistics of most common cancers worldwide in 2008 amongst females (Boyle and Levin, 2008)
3
1.2 Age specific breast cancer incidences per 100,000 population in Peninsular Malaysia 2003-2005 (Lim et al., 2008)
4
1.3 International comparisons ― Age-standardized incidences of breast cancer per 100,000 population (Lim et al., 2008)
6
1.4 Overview of the proposed framework for classification of malignant and benign abnormalities in digital mammograms
16
2.1 Anatomy and structure of the female breast 23
2.2 Microcalcification clusters (MCCs) in a breast tissue 25
2.3 Mass (lesion) in a breast tissue
26
2.4 Examples of abnormal regions of mammograms (a) Microcalcifications (b) Circumscribed mass
27
2.5 Examples of abnormal regions of mammograms (a) Spiculated mass (b) A mass classified as miscellaneous
27
2.6 Appearance of breast lesions (a) Characteristic example of a benign mass lesion (b) A benign mass which presents as a malignant lesion
29
2.7 Appearance of breast lesions (a) An example of ductal carcinoma in situ (DCIS) (b) An infiltrative ductal malignant cancer with characteristic ill-defined and spiculated borders (invasive breast cancer)
31
2.8 Benign and malignant tumors. Tumors in (a) and (b) analyzed by Rangayyan et al. (1997). Tumors in (c) and (d) analyzed by Guliato et al. (2006)
33
2.9 Examples of the most common signs of malignant abnormalities (a) Circumscribed lesion (b) Stellate lesion (c) Architectural distortion
38
2.10 The two standard views of the breast used in screening mammography (a) The craniocaudal view, from the head down (CC view) (b) A mediolateral oblique view, with the breast viewed from the side (MLO view)
48
xiv
2.11 Mammogram decomposition (a) Original mammogram image (b) Attempted decomposition of mammogram in (a) into separate breast regions
49
2.12 Magnified views showing breast nipple (a) Nipple in the breast profile (b) Breast profile without the nipple
50
2.13 Near-skin tissue of a mammogram barely visible in mammogram (a) Original mammogram image (b) Enhanced image indicating the near-skin tissue as the mask of (a)
51
2.14 Three mammograms images with different breast tissue densities. From left to right: fatty, fatty-glandular and dense-glandular
52
2.15 BI-RADS mass descriptors for (a) shape (b) margin
53
3.1 General framework for a computerized breast cancer detection system (Hutt, 1996)
70
3.2 Digital mammogram (a) Original mammogram (b) Segmentation of the mammogram into the breast area (grey) including the pectoral muscle (white) and background region (black)
72
3.3 An ROC curve ― FPF vs. TPF
83
3.4 ROC curves (a) Likelihood of a tumor being benign relative to malignant and (b) their ROC curves (Veropoulos, 2001)
84
3.5 Coordinate convention used to represent digital images (Gonzalez and Woods, 2002)
86
3.6 Grayscale image (a) A dummy matrix A (b) Image corresponding to matrix A
87
3.7 A dummy matrix �
88
3.8 Grayscale image. (a) Grayscale image corresponding to matrix B in Figure 3.7 (b) Image histogram of grayscale image in Figure 3.8(a)
90
3.9 Illustration effects of histogram equalization (Gonzalez and Woods, 2002)
92
3.10 Segmentation using thresholding techniques. (a) Image from the MIAS database (Suckling et al., 1994). Notice that the pectoral muscle has been removed to show the effects of thresholding on glandular tissue only (b) Thresholded image (Figure 3.10(a)) with a threshold value of 165
93
xv
3.11 Example of an ideal edge and a blurred edge (Gonzalez and Woods, 2002)
95
3.12 An example of dilation of set A by set B.
103
3.13 An example of erosion of set A by set B.
105
3.14 Textures (a) An image consisting of sixteen different textured regions (b) Texture segmentation of (a) produced by an automatic procedure
106
3.15 Spatial relationships of pixels defined by offsets, where � is the distance from the pixel of interest
118
3.16 Process used to create GLCMs
118
4.1 Supervised learning scheme 130
4.2 Feature extraction. The classification problem is more easily separable using the pair of features ��and �� (right) than using �� and �� (left)
131
4.3 Holdout method 134
4.4 �-fold cross-validation method 135
4.5 Leave-one-out method 136
4.6 Overfitting phenomenon. The more complex function obtains a smaller training error than the linear function (left). But only with a larger data set it is possible to decide whether the more complex function really performs better (middle) or overfits (right)
139
4.7 Schematic illustration of the bound in equation (4.11). The dotted line represents the empirical risk �� ��. The dashed line represents the confidence term. The condintuous line represents the expected risk ��. The best solution is found by choosing the optimal tradeoff between the confidence term and the empirical risk �� ��
141
4.8 A hyperplane separating different patterns. The margin is the minimal distance between the pattern and the hyperplane, thus here the dashed lines
143
4.9 Non-linearly separable patterns in two-dimensions (left). By remapping them in a three dimensional space of the second order monomials (right) a linear hyperplane separating those patterns can be found
150
4.10 General architecture of a Back-Propagation Neural Network (BPNN)
156
xvi
5.1 Flowchart of the proposed computerized breast cancer detection framework
171
5.2 Flowchart of the research framework
173
5.3 Mammography images acquired from UMMC (Dept. of Biomedical Imaging at UMMC, 2010)
176
5.4 Mammography images acquired from MIAS (Suckling et al., 2004)
177
5.5 Ground Truth (GT) markings by expert radiologists on acquired mammography datasets
178
5.6 Digitization noises (lines) in mammographic images 180
5.7 Mammogram images after noise removal using 2D median filtering
181
5.8 Histograms after applying global thresholding using � = 18 (a) Original histogram of mammogram image in Figure 5.7(a). (b) Histogram of mammogram image in Figure 5.7(a) after thresholding. (c) Original histogram of mammogram image in Figure 5.7(d). (d) Histogram of mammogram image in Figure 5.7(d) after thresholding.
184
5.9 Thresholding for separation of breast profile region from the background region. (a) Original mammogram image. (b) Mammogram image in Figure 5.9(a) after breast profile separation. (c) Original mammogram image. (d) Mammogram image in Figure 5.9(c) after breast profile separation
185
5.10 Flat disk-shaped morphological structuring element (STREL).
186
5.11 Suppression of radiopaque artifacts (a) Original grayscale image with artifact and label. (b) Thresholded image using a value of T = 18 ������ = 0.0706�. (c) Selection of the largest object with respect to Area. (d) Grayscale image with radiopaque artifacts suppressed
187
5.12 Segmented breast profile region (a) Binary image of thresholded mammogram (b) Grayscale image after background and artifact suppression
188
5.13 Cropping breast profile in mammogram images to image borders. (a) Binary image with right oriented breast. (b) Binary image in Figure 5.13(a) cropped from the left and right. (c) Binary image in Figure 5.13(a) cropped from the top and bottom. (d) Binary image with left oriented breast. (b) Binary image in Figure 5.13(d) cropped from the left and right. (e) Binary image in Figure 5.13(d) cropped from the top and bottom
190
xvii
5.14 Contrast enhancement of a mammogram image. (a) Mammogram image obtained after Image Preprocessing in Section 5.3.1.2. (b) Histogram of the original image in (a). (c) Contrast enhancement applied to the mammogram image in (a). (d) Histogram of contrast enhanced image in (c)
194
5.15 Segmentation of pectoral muscle using Seeded Region Growing. (a) Contrast enhanced breast profile right-orientated (b) Binary image of Figure 5.15(a) showing separated breast profile. (c) Segmented pectoral muscle of Figure 5.15(a) using SRG. (d) Contrast enhanced breast profile left-orientated. (e) Binary image of Figure 5.15(d) showing separated breast profile. (f) Segmented pectoral muscle of Figure 5.15(d) using SRG
197
5.16 Suppression of pectoral muscle from breast profile region (a) Subtraction of binary image Figure 5.15(c) from binary image Figure 5.15(b). (b) Subtraction of binary image Figure 5.15(e) from binary image Figure 5.15(f)
199
5.17 Binary images of the segmented breast profile. (a) Right-oriented breast profile after morphological operations. (b) Left-oriented breast profile after morphological operations
199
5.18 Pectoral muscle viewed as a right angle triangle (a) Pectoral muscle in Figure 5.15(c) viewed as a right angle triangle for a right-oriented breast profile. (b) Pectoral muscle in Figure 5.15(f) viewed as a right angle triangle for a left-oriented breast profile
201
5.19 Binary images after pectoral muscle boundary straightening. (a) Right-oriented breast profile applying straight line. (b) Left-oriented breast profile applying straight line
205
5.20 Pectoral muscle segmentation from mammogram. (a) Pectoral muscle segmentation of right-oriented breast profile. (b) Pectoral muscle segmentation of left-oriented breast profile
206
5.21 Grayscale image after pectoral muscle segmentation (a) Grayscale image of right-orientated breast profile. (b) Image histogram of Figure 5.21(a). (c) Grayscale image of left-orientated breast profile. (d) Image histogram of Figure 5.21(c)
207
5.22 Extraction of samples using different “square” ROI sizes
209
5.23 ROIs of benign abnormalities (from labeled GT data) (a) Calcification (b) Circumscribed mass (c) Spiculated mass (d) Ill-defined mass
210
5.24 ROIs of malignant abnormalities (a) Calcification (b) Circumscribed mass (c) Spiculated mass (d) Ill-defined mass (e) Architectural distortion (f) Asymmetrical mass
211
xviii
5.25 ROIs containing un-segmented background region (from labeled GT data) (a) Benign ROI. (b) Benign ROI. (c) Malignant ROI. (d) Malignant ROI. (e) Malignant ROI
213
5.26 ROIs in Figure 5.25 with segmented background region (a) Benign ROI. (b) Benign ROI. (c) Malignant ROI. (d) Malignant ROI. (e) Malignant ROI
214
5.27 The SVM training engine proposed for constructing the classification engine and performing hyperparameter optimization
224
5.28 Grid Search for selection of optimal of SVM hyperparameters ��, ��
226
5.29 SVM classification engine ‒ Trained model 228
5.30 Separating boundaries of the SVM classification engine in Figure 5.29
229
5.31 SVM testing and classification results using LIBSVM in MATLAB
231
5.32 Decision-logic system for reduction of false positives (FPs) 233
6.1 Image segmentation performance indices: TP, FP and FN 235
6.2 Binary classification confusion matrix 247
6.3 Confusion matrix after SVM testing (malignant is the + ! classand benign is the − ! class)
248
6.4 ROC curve of SVM classifier for testing with 70 samples (malignant is the + ! classand benign is the − ! class)
250
6.5 Confusion matrix after implementation of decision-logic system (malignant is the + ! classand benign is the − ! class)
251
6.6 Log-sigmoid transfer function 254
6.7 ROC curve of BPNN classifier for testing with 70 samples (malignant is the + ! class and benign is the − ! class)
257
6.8 ROC curve of OS-ELM classifier for testing with 70 samples (malignant is the + ! class and benign is the − ! class)
258
6.9 ROC curves indicating the performance of the compared machine learning techniques
259
7.1 Intelligent classification system 274
xix
LIST OF TABLES
Table No. Page
2.1 BI-RADS assessment categories (BI-RADS, 2010)
55
2.2 Elements of mammogram reporting from BI-RADS 56
3.1 Relation between, TP, TN, FP and FN ― Confusion matrix 80
3.2 Binary classification performance measures 81
3.3 Classification of the state-of-the-art texture analysis techniques
114
3.4 Standard GLCM texture descriptors (Haralick, 1973) 121
4.1 Non-linear kernels commonly used to perform a dot product in a mapped feature space in the SVM formulation
151
4.2 Summarized procedure of the SMO algorithm 154
5.1 Mammography data acquired from UMMC and MIAS database 175
5.2 ROIs extracted from acquired mammography datasets in Table 5.1
212
5.3 GLCM texture descriptors from Clausi (2002) 214
5.4 GLCM texture descriptors from Soh & Tsatsoulis (1999) 215
5.5 GLCM texture descriptors from the MATLAB Image Processing Toolbox
215
5.6 GLCM texture features calculated for each ROI sample 218
5.7 GLCM texture descriptors used to select the optimum subset of 1056 features
219
5.8 Ratio of samples used for training and testing from the UMMC and MIAS datasets
222
6.1 Comparison of classification accuracy using different ROI sizes 246
6.2 Binary classification performance metrics using the SVM as the learning machine
249
6.3 Optimum parameters for the BPNN modeling 255
6.4 Comparison of the developed framework using different machine learning techniques
256
xx
LIST OF ABBREVIATIONS
ACR American Cancer Society
ALOE Analysis of Local Oriented Edges
ANN Artificial Neural Network
ART Adaptive Resonance Theory
ASR Age-Standardized Rate
AUC Area Under Curve
BI-RADS Breast Imaging Reporting and Data System
BP Back-Propagation
BPNN Back-Propagation Neural Network
CM Completeness
CNN Convolutional Neural Network
CR Correctness
CV Cross-Validation
DCIS Ductal Carcinoma In Situ
DDSM Digital Database for Screening Mammography
DoG Difference of Gaussian
ELM Extreme Learning Machine
EU25 European Union
FFDM Full Field Digital Mammography
FIS Fuzzy Inference System
FN False Negative
FP False Positive
FPF False Positive Fraction
GA Genetic Algorithm
GLCM Gray Level Co-occurrence Matrix
GLDM Gray Level Difference Method
GLRLM Gray Level Run Length Matrix
GT Ground Truth
HIP Health Insurance Plan
KA Kernel Adatron
KKT Karush-Kuhn-Tucker
LBP Local Binary Pattern
xxi
LCIS Lobular Carcinoma In Situ
LDA Linear Discriminant Axis
LoG Laplacian of the Gaussian
LIBSVM Library for Support Vector Machine
MATLAB Matrix Laboratory
MCC Microcalcification Cluster
MIAS Mammographic Image Analysis Society
MLP Multi-Layered Perceptron
MRI Magnetic Resonance Imaging
MS-DOS Microsoft Disk Operating System
MSE Mean Square Error
NCR National Cancer Registry
OS-ELM Online-Sequential Extreme Learning Machine
PCA Principle Component Axis
PPV Positive Predictive Value
QP Quadratic Programming
RBF Radial Basis Function
RF Random Forest
RFE Recursive Feature Elimination
ROC Receiver Operating Characteristic
ROI Region Of Interest
RLS Recursive Least-Square
RMSE Root Mean Square Error
SD Standard Deviation
SGLDM Spatial Gray Level Dependence Method
SLFN Single Layer Feed-forward Neural Network
SMO Sequential Minimal Optimization
SOM Self-Organizing Map
SRG Seeded Region Growing
SRM Structural Risk Minimization
SSL Sequentially Sorted List
STREL Structuring Element
SVC Support Vector Classification
SVM Support Vector Machine
SV Support Vector
xxii
TN True Negative
TP True Positive
TPF True Positive Fraction
UCI University of California Irvine
UMMC University Malaya Medical Centre
US Ultrasonography
USD United States Dollar
VC-dimension Vapnik―Chervonenkis dimension
1
CHAPTER 1
INTRODUCTION
The discovery of a lump in the breast is one of the most frightening and feared health
problems women can face. This is due to the fact that breast cancer is the most
common cancer to afflict women in most parts of the world (Boyle & Levin, 2008).
The aim of this research is the development of a reliable tool for the detection of
breast cancer using digital mammography images. Image processing, data mining
and machine learning techniques constitute the proposed framework of this thesis.
The initial sections in this Chapter give an overview of breast cancer and the
problems and challenges faced in breast cancer detection. In the later sections of
this chapter the research motivation, objectives, contributions and an outline of
the proposed framework in this thesis is presented. Finally, the structure of the
rest of this thesis is illustrated in Section 1.6.
1.0 Overview
Breast cancer is the major cause of fatality among all cancers for women aged
between 35 to 54 years (Verma & Zakos, 2001) and continues to be the leading
cause of non-preventable cancer deaths (Kopans, 1989). Breast cancer is a serious
problem in the United States, the incidence of which continues to rise (Bassett et
al., 1997). A study made by the American Cancer Society (ACR) in 2003 estimated
that in the United States between 1 in 8 and 1 in 12 women develop breast cancer
during their lifetime (American Cancer Society, 2003a). According to these
2
statistics, on average, every 15 minutes five women are diagnosed with breast
cancer, and one woman dies of this disease (Basett et al., 1997).
According to statistics by the ACR, between 1973 and 1999, breast cancer
incidence rates have increased by approximately 40 percent (American Cancer
Society, 2003a). However, between 1989 to 1995 breast cancer mortality rates
declined by 1.4 percent per year and by 3.2 percent afterwards. These declines
have been attributed in large part, to early detection (American Cancer Society,
2003b). In the year 2009, the ACR estimated 192,370 new cases of invasive breast
cancer amongst women, as well as 62,280 cases of in situ breast cancer. Moreover,
the ACR estimated that in 2009, approximately 40,170 women were estimated to
die from breast cancer (American Cancer Society, 2009). Figure 1.1 indicates the
incidence and mortality statistics of the most prevalent cancers worldwide in 2008
amongst females (Boyle & Levin, 2008), of which breast cancer has one of the
highest statistics as compared to other cancers amongst females.
Breast cancer is the major cause of death for women in Europe (Ferlay et al.,
2007). In Europe, breast cancer represents 19 percent of cancer deaths and 24
percent of all cancer cases (Esteve et al., 1993). In 2006 in Europe the most
common form of cancer diagnosed amongst females was breast cancer, with
429,900 cases (28.9 percent of all cancer cases) and 131,900 cancer deaths. In the
European Union (EU25), breast cancer is the most common cancer with 319,900
cases (30.9 percent of all incident cases). In women, breast cancer is the major
cause of mortality with 85,300 cases (Ferlay et al., 2007).
3
Figure 1.1: Incident and mortality statistics of most common cancers worldwide
in 2008 amongst females (Boyle and Levin, 2008)
A study made on 100,000 women from 1995 to 1998 in the European Union,
indicated that for breast cancer approximately 39 deaths per year occur regardless
of the age against 40 deaths per year from 1985 to 1989, showing a change in rate
equal to -2.1 percent. The favourable trend is due to the advancements, screenings,
and the diagnosis of cancer at early stages (Levi et al. 2003). The chances of
survival significantly grow if the illness is detected at an early stage. A reduction in
breast cancer mortality for European countries in the 1990s was reported by
several research groups such as Levi et al., (2005) and Tyczynski et al. (2004).
These declines in the mortality rate have been attributed to the combined effect of
early detection of cancer and improving treatment. With the introduction of digital
mammography screening programmes throughout Europe the reduction in breast
cancer mortality (IARC Handbooks of Cancer Prevention, 2002) is expected to
decrease.
4
In Malaysia, the most common type of cancer diagnosed is breast cancer (18
percent of all cancer cases). Breast cancer is the most common type of cancer
among Malaysian women. In a 3 year period from 2003 to 2005, a total of 11,952
new cases were reported to the National Cancer Registry (NCR), Malaysia. Breast
cancer accounted for 31.3 percent of the total number of new cases in women, with
a similar percentage in each of the major ethnic groups; Malays (33.6 percent),
Chinese (30.6 percent) and Indians (31.2 percent). The age-standardized rate
(ASR) for females was 47 per 100,000 women (Lim et al., 2008). Figure 1.2
indicates the age specific cancer incidences per 100,000 population in Peninsular
Malaysia from 2007 to 2008.
Figure 1.2: Age specific breast cancer incidences per 100,000 population in
Peninsular Malaysia 2007 to 2008 (Lim et al., 2008)
The NCR in Malaysia in a report by Lim et al., 2008 estimated that the peak
incidence of breast cancer occurred in women between the ages of 50 to 60 years,
except in Indian women where the incidence peaked after the age of 60 years. The
5
incidence of breast cancer in Chinese women (ASR of 59.9 per 100,000 women)
was higher compared to Malay women (ASR of 34.9) and Indian women (ASR of
54.2). Moreover, in the same report, it was noted that the age-standardized
incidences amongst Malaysian women was lower as compared to several Western
countries: USA (92.1), Canada (78.5), England (74.4), South Australia (80.8),
Netherlands (85.6) and Denmark (81.3), but higher compared to some Asian
countries such as: Beijing (24.6), Hiroshima (36.6), Chennai (23.9) and Seoul
(20.8). Figure 1.3 provides international comparisons for age-standardized
incidences of breast cancer per 100,000 women from 2007 to 2008.
Survival through breast cancer is stage-dependent and the best survival is
observed when the cancer is diagnosed at an early stage. Mammography is
currently the best technique for reliable detection of early, potentially curable
breast cancer (Cardenosa, 1996), because it can detect cancerous cells such as:
mass lesions, microcalcification clusters (MCCS) and other suspicious anomalies
up to two years before they are palpable. Mammography has proven to be useful
in detecting cancerous cells that may be unnoticeable by physical examination
(Palmer et al. 2003).
As breast cancer incidents have increased during the recent decades, breast cancer
mortality has gradually reduced for women of all ages (Sickles, 1997). This trend
in mortality reduction is due to the adoption of mammography screening (Sickles,
1997), (Anttinen et al., 1993), (De Koning et al., 1995), (Hendee et al., 1999),
(Tabar et al., 1985), (Thurfjell et al., 1994), which allows the detection of cancer at
the early stages and the improvements made in the treatment of breast cancer
(Buseman et al., 2003).
6
Figure 1.3: International comparisons ― Age-standardized incidences of breast cancer per 100,000 population (Lim et al., 2008)
Early stage breast cancers are associated with high survival rates. Thus, the key to
surviving breast cancer is early detection and treatment. According to the ACS,
when breast cancer is confined, the five-year survival rate is almost 100 percent.
Breast cancer screening has shown to reduce breast cancer mortality. Currently,
63 percent of breast cancers are diagnosed at a localized stage, for which the five-
year survival rate is 97 percent (American Cancer Society, 2003b). The high
survival rates are attributed to the proper utilization of mammography screening
as well as high levels of awareness of the disease symptoms in the population.
1.1 Problems and Motivation
Digital mammography is a relatively new technique for the early detection of
breast cancer. It is based on accumulated density of tissues, that is, to detect
shadows. This is the reason as to why digital mammography has been considered
as an efficient tool for the detection of masses and MCCs (Khuzi et al., 2009),
7
(Verma & Zakos, 2001), (Jiang et al., 1999). The potential advantages of digital
mammography are:
1. Image Processing Capability
Clinical images presented optimally in digital format. (Appropriate contrast,
density and edge enhancement can be achieved by separating the functions
of image display and image recording).
2. Possibility of Computer-Aided Detection Methods
Several tools for computer-aided detection and computer-aided diagnosis
are under development. In the near future, these tools will provide cost-
effective mammography screening, where double reading for radiologists’
will be recommended.
3. Picture Archiving and Management
The currently technology based on a storage phosphor system, is a sub-
optimal technique because of low spatial resolution. In the near future, high
quality digital mammography using full field digital technology will be
available.
The major problem associated with mammogram screening programmes is the
large percentage of missed cancers. Studies show that during screening,
radiologists fail to detect 15 percent of breast cancers that are visible in
retrospective studies (Goergen et al. 1997), (Bird et al. 1992). Moreover, when
minimal signs are taken into account, estimates of missed cases increase to 50
8
percent (Timp et al., 2004), due to errors of perception. Eye tracker studies have
classified these mammogram screening errors into three main categories:
1. Searching Error: In this case the radiologist overlooks the abnormality. Eye
tracker experiments show that foveal sight never reached the lesion.
2. Detecting Error: The lesion has been seen by the radiologist, but the visual
dwell time was shorter than a certain threshold, for instance one second.
3. Interpretation Error: These lesions are consciously evaluated by the
radiologist, but acted on with an inappropriate decision.
Without considering recorded eye movements, search and detection errors are
those that occur when radiologists do not report the presence of a visible cancer
and interpretation errors as those that occur when the tumor is reported but not
considered actionable.
In order to identify if a breast tissue is cancerous, a biopsy is usually performed.
During a biopsy the suspicious breast tissue is removed from the patient for a
diagnostic examination. The breast tissue is removed by a surgical excision and is
diagnosed by the pathologist. An abnormality, once detected in a breast tissue, can
be classified as either benign (not cancerous) or malignant (cancerous) (Hadjiiski
et al., 1999), (Wei et al., 2005).
Although digital mammography has proven to be an efficient tool for detecting
breast cancer, the interpretation of mammograms however requires the skills and
9
experience of expert radiologists (Khuzi et al., 2009). Clinical studies have shown
that the Positive Predictive Value (PPV) (ratio of the total breast cancers found to
the total number of biopsies) is only 15 to 30 percent (Kopans, 1992), (Adler &
Helvie, 1992), (Moskowitz, 1989), (Giger and MacMahon, 1996). The performance
of a medical diagnostic test is typically measured using the Receiver Operating
Characteristic (ROC) curve analysis as presented in Section 3.3.6, where the four
performance measures: true positive (TP), false positive (FP), true negative (TN)
and false negative (FN) measure the sensitivity and specificity of the tested samples
with malignant samples being the #$%&'& !�+ !� class and benign samples being
the )!*+'& !�− !� class. Poor mammographic image quality, physician eye
fatigue, subtle nature of radiographic findings and other sources may cause
incorrect identification (misclassification) of a malignant abnormality as benign,
generally referred to as a false positive (FP) (Kocur et al., 1996), (Fogel et al.,
1998). A diagnostic model with this problem results in a bias when implemented in
the diagnosis of breast cancer. Thus, the misclassification of a benign/malignant
patient is defined as Type I/II error respectively (Fisher, 1936).
Mass lesions (or masses) and microcalcification clusters (MCCs) are the two most
important radiographic indications related to breast cancer, as they are present in
30 to 50 percent of all cancers found mammographically (Sickles 1984), (Sickles,
1986). The detection of mass lesions is a challenging task, because:
• Lesions are normally hidden or found in the dense glandular area of the
breast tissue, which is sometimes difficult to distinguish due to the variation
in shape, size and dimension (Wei et al. 2005).
10
• Lesions are usually indistinguishable from the surrounding tissue because
their features (heuristics) can be obscured as they are similar to the normal
inhomogeneous breast tissues (Bozek et al., 2008).
Similarly the detection of MCCs is also challenging, because:
• MCCs are calcium deposits of very small dimension and appear as a group of
granular bright spots in a mammogram (Wei et al., 2005). They appear as
tiny circular objects, which can be described as irregular, granular or linear
and can vary in size form 0.1mm to 1mm having an average diameter about
0.3mm (Diyana et al., 2002). Small MCCs ranging from 0.1 to 0.2mm can
hardly be seen on the mammogram due to their superimposition on the
breast parenchymal texture and noise (Diyana et al., 2003a).
• MCCs often appear in an inhomogeneous background describing the
structure of a breast tissue. Some parts of the background have features
such that the dense tissues may be brighter than the MCCs in the fatty tissue
(Diyana et al., 2003b). This is due to a large amount of absorbent tissue
(mainly fibroglandular tissue) in the dense breast image, so the image
contrast needs to be decreased.
• Some MCCs have low contrast compared to their background such as breast
tissues, blood vessels, mammary glands and fat. In other words, the
intensity and size of the MCCs can be very close to noise or an
inhomogeneous background (Diyana et al., 2002).
Thus, the masses and MCCs are relatively difficult to detect and can be overlooked
by radiologists in mammography screening. Considering the traumatic nature, cost
11
of biopsy and the relatively difficult task for radiologists to interpret
mammograms, it is desirable to develop computer-based methods which can
accurately distinguish between benign and malignant abnormalities (Mudigonda et
al., 2000). In addition, it is important to increase the positive predictive value
(PPV) without reducing the sensitivity of breast cancer detection. The use of
double reading by two or more radiologists has shown to improve the sensitivity,
but it also increased the cost of the mammogram screening process (Khuzi et al.,
2009), (Mousa et al., 2005).
In order to improve the biopsy yield ratio, computer-aided methods are desirable
for the detection of masses and MCCs and for the further classification of the
detected abnormality as benign or malignant (Mudigonda et al., 2001). Such
methods provide ease in performing initial mammogram screening and second
reading of mammograms to help radiologists in analyzing difficult cases, especially
in deciding on biopsy and follow-up recommendations (Mudigonda et al., 2000).
Using a computer-aided detection scheme, radiologists can incorporate the output
from the computer into their decision. Several recent studies have shown that
computer-aided detection improves the radiologists’ ability in differentiating
between benign and malignant abnormalities (Giger, 1999), (Jiang et al., 1996),
(Jiang et al, 1999), (Wu et al., 1993), (Huo et al., 2000), (D’Orsi et al., 1992), (Baker
et al., 1996), (Chan et al., 1999).
Computer-aided detection methods are a combination of image processing and
machine learning techniques (Hutt, 1996). Mammogram processing typically uses
image processing techniques for the purpose of suppressing artifacts and labels,
eliminating digitization noises and enhancing mammograms for optimal viewing,
12
mainly for detection of mass lesions and MCCs (Mutihac et al., 1998), (Strickland et
al., 1996), (Cernadas et al., 1996), as presented in Section 3.3 of this thesis. Many
studies provide evidence that radiologists perform better on computer enhanced
images (Aylward et al., 1998) (Netsch et al., 1998) to transform mammograms in
such a way that they can be printed or examined on a monitor optimally
(Karssemeijer & te Brake, 1996), (Bynd et al., 1997).
Machine learning techniques are typically applied in computer-aided detection
schemes for the purpose of pattern classification (Hutt, 1996). During
classification, features estimated from the Region of Interest (ROI) are used for the
training and testing (validating) a learning machine, as presented in Chapter 4. A
classifier trained on known abnormalities (mass lesions and MCCs) combines the
selected features and uses confidence measures to indicate if the tested sample is
malignant or benign. Several classification techniques have been investigated for
the detection of benign and malignant abnormalities in mammograms during the
last decade, as reviewed in Section 3.2. These classification techniques include:
Support vector machines (SVMs) (Wei et al., 2005), Artificial Neural Networks
(ANNs) (Wu et al., 1992), (Chan et al., 1995a), (Jiang et al., 1996), (Sahiner et al.,
1996), (Chan et al., 1997), (Huo et al., 1998), (Papadopoulos et al., 2002), (Zhang et
al., 2005), Linear Discriminant Analysis (LDA) (Chan et al., 1995b), (Zhang et al.,
2005), Convolutional Neural Networks (CNNs) (Sahiner et al., 1996) and the k-
Nearest neighbor (Veldkamp et al., 2000). Other machine learning techniques
include the use of statistical-based models and the general framework of Bayesian
image analysis, which was developed by Karssemeijer (1993).
13
1.2 Research Objectives and Scope
The aim of this research is to work out on the second potential advantage of digital
mammography discussed in Section 1.1, that is, computer-aided detection. This
research focuses on developing a framework of algorithms using image processing
and machine learning techniques for the detection of malignant and benign
abnormalities in digital mammography images. Prior to that, related techniques for
image processing and machine learning will be reviewed in order to identify the
most suitable approach for the detection of malignant and benign abnormalities,
which includes the detection of mass lesions and MCCs.
The goal of this research is to increase the diagnostic accuracy of image processing
and machine learning techniques for optimum classification between malignant
and benign abnormalities as well as to the reproducibility of mammographic
interpretation. In order to achieve the goal of this research, the following research
objectives are set:
1. To investigate and apply existing image processing algorithms and machine
learning techniques in order to detect mass lesions and MCCs in digital
mammograms.
2. To develop a system using the investigated techniques, for the classification
of malignant and benign abnormalities, using a combination of image
processing algorithms and machine learning techniques.
14
3. To apply the identified algorithms and techniques in order to reduce the
number of misclassified malignant cancers, namely, false positives (FPs)
(see Section 3.3.6).
4. To verify the reliability and accuracy of the developed system using the
datasets from different sources.
5. To perform a comparative study for identifying the most suitable machine
learning technique for the classification of benign and malignant
abnormalities.
The proposed framework in this research relies mainly on image processing
algorithms and machine learning techniques. For a deeper understanding of the
proposed framework refer to Section 1.4.
1.3 Research Significance and Contribution
In addition to the problems discussed in Section 1.1 for the detection of masses
and MCCs, in digital mammograms, it is observed that most mass lesions have a
stellate appearance in mammograms. Moreover, their central masses are typically
irregular with ill-defined borders and can vary in size from a few millimetres to
several centimetres in diameter. Due to these problems and the problems
presented in Section 1.1, the detection of malignant and benign masses and the
detection of MCCs has become a challenging task.
Applying a combination of image processing and machine learning techniques for
the detection of abnormalities in digital mammograms requires feature
15
calculation/estimation from the Region of Interest (ROI), namely, the abnormal
region. In general, it is difficult to determine the size of the neighbourhood (pixels)
or the ROI that should be used to calculate the relevant features from the abnormal
regions (masses and/or MCCs). If the size of the ROI is too large, small masses
and/or MCCs may be missed, while if the size of the ROI is too small, parts of large
masses and/or MCCs may be missed. This poses a challenging task in the detection
of breast cancer. Thus, the primary contribution of this research is:
• To determine the most suitable ROI (neighbourhood) size of mass lesions
and MCCs for the purpose of feature computation (extraction).
This specifically addresses the difficulty of predeterming the ROI size for the
purpose of feature estimation (extraction). The detection of masses and/or MCCs is
regarded as one of the hardest to solve in the field of recognition of objects into
images (see Section 1.1). The difficulty in carrying out research, such as this, lies
not only in the process itself, because even radiologists find it challenging to
identify masses and MCCs given their variability in shape, size and dimension.
Thus, the secondary contribution of this research is:
• To demonstrate that advanced machine learning techniques, namely,
Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs)
can effectively solve pattern classification problems.
The secondary contribution of this research provides the basis for conducting a
comparative analysis between different machine learning technologies, such as
ANNs and SVMs which complies with the fifth research objective in Section 1.2.
16
1.4 Research Methodology and Proposed Approach
The literature discussed in Chapters 2, 3 and 4 of this thesis presents the state-of-
the-art image processing and machine learning algorithms applied for the
detection of malignant and benign abnormalities in digital mammograms. Figure
1.4 illustrates an overview of the basic framework developed in this research. For a
detailed illustration of the proposed framework and applied techniques, refer to
Figure 3.1 and Figure 5.1. The inputs into the proposed framework are digital
mammogram images, whereas the output of the system indicates that the
abnormality in the ROI of the input image is malignant or benign.
Figure 1.4: Overview of the proposed framework for classification of malignant
and benign abnormalities in digital mammograms
The literature reviewed in Chapter 2 and Section 3.3 indicates that digital image
processing techniques need to be applied to the mammogram images for the
purpose of noise removal and pectoral muscle segmentation. Usually, mammogram
preprocessing techniques includes: noise removal, background suppression and
artifact/wedge and label removal as discussed in Section 3.3.1. Since the breast
profile needs to be optimally segmented from the background region, the pectoral
Image Processing
Machine Learning
Region of Interest (ROI) Selection
Image Processing and Image Segmentation
Texture Feature Estimation and
Selection
SVM Classification Engine (Training)
SVM Model Parameter
Optimization
SVM Classification (Testing and Validation)
Input:
Result:
Digital Mammogram
Malignant or Benign
17
muscle is also suppressed from the mammograms since it may bias the procedures
in the detection of malignant and benign abnormalities as discussed in Section
3.3.2. The ROIs (benign and malignant samples with mass lesions and/or MCCs)
are extracted from the segmented mammogram images using the Ground Truth
(GT) markings from the radiologists’ interpretation.
Digital mammograms are known to contain prominent textural information
regarding the shapes and sizes of masses and MCCs. Texture based features are
known to increase the performance of machine learning algorithms such as ANNs
and SVMs (Belotti et al., 2006), (Makinacı et al., 2005), (Jirari et al., 2005). The
approach proposed in this research uses texture features for the purpose of
pattern classification, as discussed in Section 3.3.3. Thus, texture descriptors are
computed from the ROIs representing the malignant and benign samples.
As discussed in Section 1.3, SVMs can effectively solve binary classification
problems using noisy data. The theoretical concepts of machine learning and SVMs
are discussed in Chapter 4. Binary classification results obtained from the SVM
testing in Figure 1.4 are analyzed using Receiver Operating Characteristic (ROC)
curves presented in Section 3.3.6, to estimate the classification accuracy of the SVM
for unseen samples.
1.5 Benefits of Image Processing and Machine Learning Techniques
Digital mammography leads itself well to breast cancer detection, where image
processing algorithms enable computers to indicate suspicious areas of the breast
that contain masses, MCCs or other prominent signs of breast cancer.
18
Image processing algorithms have the potential to increase the diagnostic accuracy
by reducing the number of FPs (misclassified malignant cancers), through the
application of image segmentation. Using segmentation, the unimportant parts in
the ROIs (the breast region near to the breast border and the boundary between
the segmented pectoral muscle and the breast tissue) can be eliminated, thus
making feature estimation and calculation more accurate, as discussed in Section
5.4.1.1. The benefits obtained of using the proposed framework in Figure 1.4 are as
follows:
1. The proposed system will aid radiologists in their diagnosis by indicating
suspicious abnormalities in mammograms. Thus, the system will act as a
second reader after the radiologists.
2. The proposed system will substantially reduce the number of false positives
(FPs), (see Section 3.3.6), which will eliminate the need of performing
unnecessary biopsies and save cost.
3. This system will reduce patient examination time by inspecting
mammograms and reporting the findings within a few seconds.
In breast cancer diagnosis, the weakest link has always been the radiologists, since
it is the radiologists who must find mass lesions and MCCs in order to make a
diagnosis. So, the radiologists can refer to this system for a second opinion as it is
difficult to distinguish between malignant (cancerous) from benign (non-
cancerous) tissues due to their similar nature and visual features (Veldkamp et al.,
2000), (Rahbar et al., 1999), (Jiang et al., 1998).
19
1.6 Thesis Overview
This thesis is arranged in a methodical manner. It is organized into seven chapters
comprising of this introduction chapter and six further chapters as follows.
Chapter 2 discusses breast cancer detection and digital mammography. General
information about the structure and functions of the breast, breast tumors and
literature regarding breast cancer screening is presented at first. Next, background
issues concerning different imaging modalities such as digital mammography,
ultrasonography (US) and magnetic resonance imaging (MRI) are reviewed and
discussed. Towards the end of this chapter, the analysis and interpretation of
digital mammograms using the BI-RADS lexicon is discussed.
Chapter 3 presents and discusses the background literature on the computer
processing of digital mammograms. In Section 3.1, computer-aided detection
systems are introduced with the literature review of computerized breast cancer
detection techniques presented in Section 3.2. Section 3.3 highlights and identifies
the key techniques and algorithms used in this research to develop a framework
for the computerized detection of breast cancer. Section 3.4 presents the
fundamental concepts of digital image processing with emphasis on image
segmentation techniques used in digital mammography applications. Lastly,
Section 3.5 emphasizes on the use of texture-based analysis for the purpose of
feature extraction in pattern classification problems.
In Chapter 4, an overview of pattern recognition is presented, with particular
emphasis on a specific machine learning technique, namely, the Support Vector
Machine (SVM). SVMs will be used intensively in this research. The reason for
20
using SVM as the main machine learning technique for this research is discussed in
Sections 1.3 and 3.3.5. Section 4.1 presents some introductory notions regarding
the theoretical concepts of learning machines. Section 4.2 introduces the
fundamental concepts of the statistical learning theory and presents the
mathematical formulation of the SVM developed by Vapnik (1998) which describes
the statistical aspects of automated machine learning. Towards the end of this
chapter, Section 4.3 presents the theoretical concepts of ANNs whereas Section 4.4
discusses a Recursive Feature Elimination (RFE) technique used for the selection
of the optimal subset of texture features for the learning machine (SVM).
Chapter 5 presents the modeling of the framework (system) proposed in Chapter 1
for the classification of benign and malignant abnormalities in digital
mammograms. As discussed in Chapter 3, the proposed framework is composed of
two main techniques, namely, image processing and machine learning. The image
process and machine learning techniques identified in Section 3.3, are applied in
the proposed system in this chapter, which are discussed in Section 5.1 and 5.2.
The modeling of the system consists of three main stages, namely: Mammogram
Image Processing (Section 5.3), Texture Feature Extraction and Selection (Section
5.4) and Classification Engine (Section 5.5). Sections 5.3 through 5.5 describe each
stage in detail during the development of the proposed framework.
Chapter 6 presents the experimental results of the developed system in Chapter 5.
Section 6.1 presents and discusses the SVM training results relative to the
memorization and learning of the binary SVM classifier. Section 6.1 also presents
and discusses the SVM testing and validation results for unseen samples. In order
to perform a comparative research, Section 6.2 presents the experimental results
21
obtained after evaluating the developed framework using different machine
learning algorithms other than the SVM. The experimental results of the compared
machine learning models are discussed in the last part of Chapter 6.
Chapter 7 concludes and summarizes the research contributions made. The
achievements and objectives of the research with respect to the experimental
results obtained are highlighted along with the key findings and significance of the
research. This chapter also discusses the impact and significance of the developed
system to radiologists and hospitals for mammography screening and
interpretation. Radiologists and clinicians will benefit from the developed system
as it will assist them in their diagnosis by acting as second readers.
22
CHAPTER 2
DIGITAL MAMMOGRAPHY
2.0 Overview
This chapter discusses breast cancer detection and digital mammography. General
information about the structure and functions of the breast, breast tumors and
literature regarding breast cancer screening is presented at first. Next, background
issues concerning different imaging modalities such as digital mammography,
ultrasonography (US) and magnetic resonance imaging (MRI) are reviewed and
discussed. Towards the end of this chapter, the analysis and interpretation of
digital mammograms using the BI-RADS lexicon is discussed.
2.1 Breast Anatomy and Cancer
The most important anatomical structures of the breast are shown in Figure 2.1.
The breast consists of two components. The first component is concerned with
milk production and is known as the epithelial component. The second component
consists of fat and connective tissue, which supports and protects the structure of
the breast (Bassett et al., 1997).
The epithelial component of the breast consists of a tree-like branching pattern of
milk ducts that come together at the nipple. The leaves of this tree are formed by
the lobules which are the secretory units of the breast. Each lobule consists of a
number of acini connecting to an intra-lobular duct. The acini are composed of two
types of cells, namely, the epithelial and myo-epithelial. The epithelial cells secrete
23
a variety of glyco-proteins and during lactation they also produce milk. The myo-
epithelial cells are capable of contracting during breastfeeding. Each intra-lobular
duct connects with an extra-lobular duct, and this together with the lobule, is
called the terminal ductal lobular unit (Chu et al., 1988).
Figure 2.1: Anatomy and structure of the female breast
The extra-lobular ducts of the breast link together and form sub-segmental ducts,
which in turn form the segmental ducts. These ducts drain milk from different
segments or lobes of the breast. In total, the breast consists of 15 to 20 lobes,
which are roughly pyramidal in shape with the apex directed towards the nipple.
The non-epithelial component of the breast consists mainly of fatty tissue. There
are no muscles in the actual breast, but there are a series of muscles behind and
underneath the breasts. These muscles work together with a ligament called
Cooper ligament to support the weight of the breasts (Chu et al., 1988).
Breasts contain lymph vessels, which are very important in fighting against
diseases in the body. Lymph is a clear fluid that contains tissue fluid, waste
24
products, as well as immune system cells. The lymph system consists of lymph
nodes and lymph vessels that transport the lymph to the lymph nodes. Most of the
lymph vessels that go through the breast carry the lymph to the lymph node
underneath the arm pit, called axillary nodes. The other lymph vessels carry the
lymph to the lymph nodes which are inside the chest, called the internal mammary
nodes, or the lymph nodes above or below the collarbone, called the
supraclavicular or infraclavicular nodes respectively. Lymph veins can also carry
possible diseases to lymph nodes which might increase the spread of the disease,
for example breast cancer (malignant) cells (Boyle, & Levin, 2008).
Breast cancer is the major cause of fatality among all cancers for women aged
between 35 to 54 years (Verma and Zakos, 2001) and continues to be the leading
cause of non-preventable cancer deaths amongst females, as indicated in Figure
1.1. Breast cancer is developed when the cells of the breast become abnormal
(malignant) and spread without order or control. The malignant cells then form a
tissue and turn into a tumor. This tumor typically grows into nearby tissues or
breaks away and enters the bloodstream or lymphatic system which can affect
other organs. The spreading of breast cancer is generally referred to as metastasis
(Kopans, 1989).
The most common and effective method for detecting breast tumors in their early
stages is by performing mammogram screening. Mammography is currently the
most effective modality used to detect tumors in the breast tissue (Sickles, 1997),
(Anttinen et al., 1993), (De Koning et al., 1995), (Hendee et al., 1999), (Tabar et al.,
1985), (Thurfjell et al., 1994) that can indicate potential clinical problems, such as
the: asymmetries between breasts, architectural distortion, confluent densities
25
associated with benign fibrosis, microcalcification clusters (MCCs) and mass
lesions. By far, the two most common features that are typically associated with
breast tumors are MCCs and mass lesions, which are discussed in the following
sections.
Figure 2.2: Microcalcification clusters (MCCs) in a breast tissue
2.1.1 Calcifications
Calcifications are small mineral (calcium) deposits within the breast tissue that
appear as localized high intensity regions in the mammogram. There are two types
of calcifications: microcalcifications and macrocalcifications. Macrocalcifications
are coarse, scattered calcium deposits. These deposits are usually associated with
benign conditions and rarely require a breast biopsy. Microcalcifications on the
other hand are isolated calcium deposits that normally appear in clusters or are
found embedded in a lesion. Individual microcalcifications typically range in size
from 0.1 to 1.0mm with an average diameter of about 0.5mm. A microcalcification
cluster (MCC) is typically defined to be at least three microcalcifications within a
1cm2 region, as shown in Figure 2.2. About 30 to 50 percent of non-palpable
cancers are initially detected due to the presence of MCCs (Feig & Yaffe, 1995).
Similarly, in a large majority of the ductal carcinoma in situ (DCIS) cancers, MCCs
are present (Monsees, 1995).
26
Figure 2.3: Mass lesion in a breast tissue
2.1.2 Mass Lesions
Breast tumor is often represented as a mass lesion with or without the presence of
MCCs (American Cancer Society, 2003a). A cyst, which is a non-cancerous
collection of fluid, may appear as a mass in the film. However, ultrasound or fine
needle aspirations can distinguish the difference. The similarity in intensities with
the normal tissue and morphology with other normal textures in the breast makes
it more difficult to detect masses compared with calcifications (Feig & Yaffe, 1995).
The location, size, shape, density, and margins of lesions are useful for the
radiologist in evaluating the likelihood of a cancer (Evans, 1995). Most benign
masses are well circumscribed, compact, and roughly circular or elliptical, as
shown in Figure 2.3. Malignant lesions usually have a blurred boundary, irregular
appearance and sometimes are surrounded by a radiating pattern of linear
spicules (Evans, 1995). However, some benign lesions may have a spiculated
appearance or blurred periphery.
Mass lesions and MCCs are abnormal regions in mammograms. Examples of
mammograms with MCCS and masses are shown in Figure 2.4 and Figure 2.5, with
the Ground Truth (GT) data (in Section 5.2.1) superimposed. The GT data for a
27
mammography dataset are the radiologist’s findings in the diagnosis of
mammographic abnormalities which are the location, size and shapes of suspicious
masses and/or MCCs found, as shown in Figure 5.5. As observed in Figures 2.4 and
2.5, there are several different lesion types and lesions can either be malignant or
benign. Malignant tissues indicate cancer, whereas benign tissues indicate non-
cancerous cells, i.e. abnormal tumors. The following section discusses cancerous
and non-cancerous breast tumors in detail.
(a) (b)
Figure 2.4: Examples of abnormal regions of mammograms
(a) Microcalcifications (b) Circumscribed mass
(a) (b)
Figure 2.5: Examples of abnormal regions of mammograms
(a) Spiculated mass (b) A mass lesion classified as miscellaneous
28
2.2 Breast Tumors
Breast tumor is identified by different names, depending on where it starts in the
woman's breast. Scientists are not sure of the exact cause of breast cancer, but they
have identified high risk factors for this disease. The most common factors include:
age, family history and personal history. The most common symptom is a painless
lump in the breast. At times, a painful lump in the breast turns out to be cancer
(malignant). One or more lumps in the armpit can be a symptom of breast cancer;
however, they can also be due to non-cancerous (benign) conditions. The bleeding
from the nipple can indicate the presence of cancer, especially if the bleeding
occurs from one breast only (Buseman et al., 2003). A more difficult symptom to be
identified is the thickening of the tissue in the breast. Any changes in the breast
size or shape can be due to cancerous (malignant) or non-cancerous (benign)
conditions. Although benign conditions in the breast can be understood by their
symptoms, biopsies need to be performed to understand if the irregularity in the
breast is non-cancerous or otherwise. Other symptoms indicating the possibility of
breast cancer are: redness of the skin over a portion of the breast, redness or
scaliness of the nipple or any nipple pain or retraction (nipple turning inward), an
orange peel appearing on the skin and dimpling of the skin (Basett et al., 1997).
In our research, we distinguish two types of breast abnormalities, namely, benign
and malignant. The following sections discuss cancerous and non-cancerous breast
tumors in detail.
29
2.2.1 Non-Cancerous Breast Tumors
Non-cancerous (benign) tumors of the breast comprise fibro-adenoma, duct
papilloma, adenoma and connective tissue tumors. The most common benign
breast tumor is the fibro-adenoma. This tumor is a combined product of both
connective tissue and epithelial cells in the breast. Most benign masses are
circumscribed due to the absence of infiltration (Palmer et al. 2003).
(a) (b)
Figure 2.6: Appearance of breast lesions
(a) Characteristic example of a benign mass lesion
(b) A benign mass which presents as a malignant lesion
Figure 2.6(a) shows a characteristic example of a benign mass lesion. The shape is
oval and the border is sharply delineated. Benign mass lesions may also be
suspicious (presented as malignant lesions), as shown in Figure 2.6(b). Based on
these reasons, mammographically it is difficult to distinguish between benign and
malignant lesions (Khuzi et al., 2009).
2.2.2 Cancerous Breast Tumors
Cancerous (malignant) breast tumors can be classified into two main types: (i)
Non-Invasive breast cancer, and (ii) Invasive breast cancer. The following sections
discuss cancerous breast tumors in detail.
30
2.2.2.1 Non-Invasive Breast Cancer
Non-Invasive―in situ―cancer consists of malignant cells that replace the normal
epithelial cells, lining the ducts or lobules in the breast tissue. These malignant
cells are confined to the basement membrane and have not yet invaded the breast
stroma or lymphatics. The two forms of non-invasive breast cancer are: (i) Ductal
Carcinoma In Situ (DCIS) and (ii) Lobular Carcinoma In Situ (LCIS) (Lu and
Bottema, 2001), which are described as follows:
• Ductal Carcinoma In Situ (DCIS): DCIS is a malignancy of the epithelial
cells lining the lactiferous ducts (usually the terminal ducts) without
penetration of the ductal basement membrane. The prognosis of untreated
DCIS is not precisely known, as most patients are treated with mastectomy.
It is estimated that about one-third to half of the untreated patients
eventually will develop invasive cancer, usually in the same quadrant of the
breast where the first lesion develops. Mammographically DCIS is often
characterized by the presence of microcalcifications. When there is
extensive fibrosis, DCIS may also present as a palpable mass (Lu and
Bottema, 2001). Figure 2.7(a) shows an example of DCIS.
• Lobular Carcinoma In Situ (LCIS): In LCIS the lobules are expanded by a
uniform population of small yet atypical cells. Usually this process
obliterates the lumen of the acini. These atypical cells do not penetrate
through the walls of the lobules. LCIS rarely gives rise to mammographic
abnormalities. It is often found in biopsies that have been done for other
reasons such as removal of benign lesions. LCIS is a risk factor for
31
developing breast cancer. The majority of patients are therefore managed
by careful follow ups (Lu and Bottema, 2001).
2.2.2.2 Invasive Breast Cancer
Invasive breast cancer, also known as infiltrating cancer, occurs when malignant
cells have spread beyond the ducts or lobules to other parts of the breast or body.
Invasive cancers vary in size from less than 10mm in diameter to over 80mm, but
are usually 20 to 30mm at presentation (Vitak, 1998).
(a) (b)
Figure 2.7: Appearance of breast lesions
(a) An example of ductal carcinoma in situ (DCIS)
(b) An infiltrative ductal malignant cancer with characteristic ill-defined and
spiculated borders (invasive breast cancer)
Ductal carcinoma accounts for about 80 percent of all invasive breast cancer cases.
These tumors are believed to arise from epithelial cells of the terminal ductal
lobular unit. It is generally thought that ductal carcinoma starts as a DCIS. Less
common types of breast cancer include: lobular carcinoma, medullary carcinoma,
32
tubular carcinoma, mucinous carcinoma, cribriform carcinoma and papillary
carcinoma (Popli, 2001). Figure 2.7(b) shows an example of an infiltrative cancer.
Breast cancers can infiltrate locally to the skin and the muscle, or metastasise to
more distant sites via lymphatics or the bloodstream. The most common spread
via lymphatics is to the axillary lymph nodes. Metastasis via the blood stream most
frequently involves the lung and the liver, but adrenals and brains are also
common sites for metastasis. When a woman has invasive breast cancer the
prognosis depends among others on the histological grade and behavioral
characteristics of the tumor and the presence of the metastatic spread (Vitak,
1998).
Considering histology, tumors can be graded based on their degree of
differentiation. Well differentiated tumors often have a better prognosis than
poorly differentiated tumors. Behavioral characteristics that influence prognosis
include the growth rate and the receptor status of a tumor. Tumors with lower cell
growth rates generally behave better. The presence of oestrogen receptors
indicates that the tumor cells have a higher degree of functional differentiation
resulting in a better prognosis (Popli, 2001). Tumor spread is also associated with
a worse prognosis than when there is no evidence of metastasis. Although these
factors may predict how individual cancers will behave, this has not led to an
improvement of patient survival. Mammography screening on the other hand has
proven to improve survival rates (Vitak, 1998), as discussed in Section 2.4.
33
2.3 Differentiating Between Breast Tumors
Although there are many types of breast abnormalities, it is possible to have a
general differentiation between benign and malignant breast tumors using their
boundary shapes with the surrounding breast tissue. This differentiation can be
performed by examining spiculations on the malignant tumor, which can easily be
observed using mammography or ultrasound techniques. Spiculation is a stellate
distortion caused by the intrusion of breast cancer into the surrounding tissue and
its existence is very important for tagging the tumor as malignant. There are many
successful techniques based on mammography or ultrasound to quantify the
degree of spiculations for a successful decision of differentiation between the
benign and malignant type of breast tumors (Huang et al., 2004). Figures 2.8(a)
and 2.8(b) indicate a benign and malignant tumor as analyzed by (Rangayyan et al.,
1997), whereas Figure 2.8(c) and 2.8(d) show a benign and malignant tumor
analyzed by (Guliato et al., 2006).
(a) Benign (b) Malignant
(c) Benign (d) Malignant
Figure 2.8: Benign and malignant tumors. Tumors in (a) and (b) analyzed by
Rangayyan et al. (1997). Tumors in (c) and (d) analyzed by Guliato et al. (2006)
34
Breast tumors and masses appear as dense regions in mammographic results and
although there might be some exceptions to the general rule, a typical benign mass
has a round, smooth, and well-circumscribed boundary, while a malignant mass
has a spiculated, rough and blurry boundary (Varela et al., 2006), (Cheng, et al.,
2006).
As it is difficult for radiologists to differentiate between benign and malignant
tumors on mammograms, many recent studies have shown that techniques can be
developed to assist radiologists to decide on the type of tumor by using
quantitative methods. Guliato et al. (2006) derived a mathematical model for
mammograms, in order to derive polygonal models of contours for an accurate
classification of tumors. Rangayyan et al. (2000) developed a method to quantify
the sharpness of the tumor boundaries. Varela et al. (2006) divided the tissue
tumor border into three sections and analyzed these sections independently to
decide if the mass lesion is benign or malignant by considering the shape of its
interface shape. Kim and Min (2002) developed a mathematical model to count the
number of jags of the breast tissue and breast tumor interface in order to
differentiate between benign and malignant tumors. The most common and
effective method for detecting breast tumors in their early stages is by performing
mammogram screening. Mammography is currently the most effective modality
used in breast cancer screening programmes (Sickles, 1997), (Anttinen et al.,
1993), (De Koning et al., 1995), (Hendee et al., 1999), (Tabar et al., 1985),
(Thurfjell et al., 1994), which is discussed in detail the following sections.
35
2.3.1 Mammography
At the current moment the modality of choice for breast cancer screening is
mammography. Mammography is an X-ray technique developed specifically for the
breast. It is based on the differential absorption of X-rays between breast tissue
components such as the fat, connective tissue, tumor tissue and calcifications.
Mammography is used both as a clinical tool to examine symptomatic patients and
for screening purposes (Sickles, 1984). Requirements for mammography are high
contrast, high spatial resolution, and minimal radiation exposure. High contrast is
needed because differences in density between normal and pathologic structures
of the breast tissue are small. The detection of MCCs requires both high contrast as
well as a high spatial resolution. Minimal radiation exposure is essential as in
screening programmes women frequently undergo mammography, often annually.
Breast cancer can be recognized mammographically by the presence of a mass
lesion (masses) or MCCs. The characteristics of mass lesions and MCCs are as
follows (Burrel et al., 1996):
• Mass Lesions
Most breast tumors, benign as well as malignant tumors are present as a
focal mass lesion. The task of radiologists therefore is to discriminate
between benign and malignant lesions. When a radiologist considers a
lesion suspicious for containing a malignancy the patient will undergo
additional examinations. The most important sign of malignancy is the
presence of spiculation, which is a stellate pattern of lines directed towards
the centre of the lesion. The border of a mass may also give information
about the potential malignancy of a lesion. Benign masses are often
characterized by sharp, circumscribed borders. Malignant masses on the
36
other hand have ill-defined or spiculated borders. The sharpness of the
border however cannot be used as solitary criterion to identify malignancy,
as some malignant masses, for example medullary carcinoma, colloid
carcinoma and intracystic carcinoma, have circumscribed borders as well
(Sickles, 1984). Moreover benign masses generally have poorly defined
margins, for instance due to overlapping of the breast tissue or fibrosis.
When a lesion is probably benign or when multiple similar masses are
found in the breast the patient is often placed in a follow up protocol.
Otherwise further examination is necessary to determine the nature of the
mass (Sickles, 1986).
• Microcalcifications
Another sign of malignancy is the presence of microcalcification clusters
(MCCs). Microcalcifications develop in microscopically small cavities inside
the lobuli or ducti. Microcalcifications inside the lobular unit are often due
to benign conditions such as adenosis or fibro-adenoma (Popli, 2001). MCCs
of ductal origin are more suspicious and may be the first sign of breast
cancer. Intra-ductal microcalcifications can be diagnosed as benign or
malignant by analyzing the shape of the cluster and the shape of the
individual microcalcifications. Studies show that irregular, pleomorphic
shapes of microcalcifications have a higher probability of being associated
with malignant disease than those with round shapes and uniform size
(Sickles, 1984).
37
2.4 Screening for Breast Cancer
Early detection of breast tumors, especially breast cancer, can save thousands of
lives each year. Screening is a very important step for early detection of diseases,
which can locate breast cancers while they are still small in size and confined to
the breast before they cause any symptoms. Screening is also important because
breast cancers that cause discomfort to patients and that are big enough to be
easily felt, tend to have already spread outside the breast to the other parts of the
body (Monsees, 1995).
During breast cancer screening, the incorrect identification of a malignant
abnormality as benign in breast cancer patients is generally referred to as a false
positive (FP) (Kocur et al., 1996), (Fogel et al., 1998), as discussed in Section 3.3.6.
The aim of breast cancer screening is early detection of breast cancers while
keeping the number of false positive (FP) detections at a minimum. Similar to
breast cancer screening, the performance of a medical diagnostic test is typically
measured using the Receiver Operating Characteristic (ROC) curve analysis (see
Section 3.3.6), where the four performance measures: true positive (TP), false
positive (FP), true negative (TN) and false negative (FN) measure the sensitivity
and specificity of the tested samples.
The earlier breast cancers are detected; the better the treatment options are
available. A high patient recall rate, i.e. the percentage of mammographically
screened women that is recalled for further assessment, generally improves the
detection rate. This however, can lead to an increase in the number of FP
detections resulting in unnecessary examinations (biopsies) and additional costs.
Most countries have patient recall rates between 3 to 5 percent.
38
In many countries breast cancer screening programs using mammography have
been started to detect cancers as early as possible. A screening program is defined
as a program where an asymptotic group is invited to examine a specific disease on
a regular basis. For breast cancer screening programs, only women are invited due
to the very low incidence rate among men. A number of parameters must be
chosen for a breast cancer screening program. The two parameters are: (i) the age
range of women that are invited and, (ii) the time interval between two screening
rounds. It is a highly debated subject at what age women should be invited for
their first screening (Peer et al., 1995), varying in practice between 40 and 50
years. Below the age of 40, the incidence rate of breast cancer is extremely small,
increasing rapidly between the age of 40 and 50, which continues to increase more
gradually for older women. The problem with screening young women is that their
breasts contain much glandular tissue, yielding mammograms that are difficult to
read due to denser tissues.
(a) (b) (c)
Figure 2.9: Examples of the most common signs of malignant abnormalities
(a) Circumscribed lesion (b) Stellate lesion (c) Architectural distortion
Breast cancers in young women are often aggressive and fast growing tumors,
requiring short intervals between two screenings. After menopause, the breast
39
becomes less dense, making successful screening for small cancers more feasible.
The upper limit of age for which women are invited for screening varies between
65 and 75 (Van Dijck et al., 1997).
If the interval between two successive screening rounds is too large, a number of
tumors that are detected in screening have already reached a stage with a lower
chance of successful treatment. Tumors occurring during this interval are known
as interval carcinomas. A large number of interval carcinomas may indicate that
the screening interval should be made shorter. A short interval period will have a
larger effect on the reduction of mortality, but is more expensive and women are
exposed to a higher number of X-ray doses. In the United Kingdom, the screening
interval is 3 years, a period that is considered too long by some researchers (Dean,
1996), whereas in Sweden and the Netherlands it is 2 years (Tabár et al., 1997).
However, in Malaysia, there is an absence of a national screening programme for
breast cancer. As, breast cancer is the commonest cancer among Malaysian
women, (18 percent of all cancer cases) (Lim et al., 2008), age dependent screening
should be performed by the NCR, i.e. 1 to 1.5 years for women below 50 years and
2 years for women over 50 years (Lim et al., 2008), (Khuzi et al., 2009).
In the United Kingdom only oblique films are used in screening, most other
countries use both oblique and cranio-caudal films. The way mammograms are
read also varies between countries. In some countries (for example the
Netherlands) mammograms are examined by two radiologists, called double
reading. Various approaches can be used to combine the findings of the two
radiologists. Thurfjell et al. (1994) found that the sensitivity increases when
double reading is practiced (when a case is recalled if either one of the radiologists
40
finds it suspicious), without changing the positive predictive value (PPV). In
medical diagnostic tests using ROC curves, sensitivity represents the ratio of
tumors which are marked and classified as tumor, to all marked tumors. Specificity
represents the ratio of tumors which are not marked and also not classified as
tumor, to all unmarked tumors, as indicated in Section 3.3.6. The study by Thurfjell
et al. (1994) was criticized by Beam & Sullivan (1994) because the demonstrated
increase in sensitivity is a mathematical fact; the question should be if the increase
in sensitivity is worth the decrease in specificity. Another study has shown that
double reading based on consensus between radiologists is a cost-effective
screening procedure (Brown et al., 1996).
Proving the efficacy of a screening program in a traditional epidemiological way is
difficult due to the lack of an effective control group (Dean, 1996). If half of the
population is offered screening, part of this group will not participate. Women in
the group that is not offered screening and subsequently develop breast cancer
typically do not seek medical attention until tumors are already in a late and
incurable stage. If the group of non-participating women is large, a serious self-
selection bias is present in the study. Even more important, women in the control
group cannot be denied to have mammography on a regular basis. Women that are
in a high risk group will especially request for mammography on a regular basis.
Thus, if these high risk women are in the control group, it would then, reduce the
number of cancers that are found at an incurable stage in the control group
creating a biased experiment. This effect is called contamination and is a serious
problem when the effect of screening is studied.
41
These factors make it difficult to prove a significant mortality reduction in
screening. Comparing the number of breast cancer deaths with the number before
screening is a common way to solve this problem, but it suffers from several
drawbacks: the incidence of breast cancer may have changed, the treatment of
breast cancer may have improved, or women may be more aware of abnormalities
and seek medical assistance earlier than they might have done before. Another
complicating factor is the long time it takes before a screening program reaches its
maximal reduction in mortality.
Early screening programs were based on breast examination using palpation,
either by the woman or a physician. No significant reduction in mortality has been
reported on randomized trials using this type of screening (Newcomb et al., 1991)
although a few other studies suggest a small benefit (Baines, 1992). In 1963 the
Health Insurance Plan (HIP) project was started in New York, the first large
screening experiment using mammography as the main screening modality,
together with palpation. A reduction in mortality was found for women in the
group that underwent screening, a success that could be achieved because
palpation and mammography were hardly practiced by women in the control
group (Shapiro et al., 1971). The success that was reported stimulated other
countries like Sweden, Finland, the United Kingdom, Canada and the Netherlands
to start experiments with breast cancer screening. The results of large numbers of
randomized studies have been published by Dean (1996). It is commonly accepted
that these studies show a reduction in mortality for women that take part in a
screening program, especially for the age group between 50 and 70 years old. This
is confirmed by other non-randomized and cohort studies in the United Kingdom
and the Netherlands.
42
In a number of studies the possible benefit of inviting young women between 40
and 50 years old to a screening program was examined, but no unequivocal results
were obtained. Some studies suggested a reduction in mortality (Thurfjell et al.,
1996); others did not find evidence for this (Peer et al., 1995). It was shown in Van
Dijck et al. (1997) that screening is beneficial at least until the age of 75. Due to the
limited number of women over 75 that participated in screening, no significant
results could be obtained for this age group.
An important trial of mammography screening was conducted between 1977 and
1984 in Sweden (Tabár et al. 1985). This trial concerned women aged 40 and
older. The women were divided randomly into two groups, namely the: study
group and control group. Each woman in the study group was offered screening
every 2 or 3 years depending on their age, while women in the control group were
not offered screening. The results obtained after seven years of follow up showed a
31 percent reduction in breast cancer mortality rate for the women in the study
group who were invited for screening (Shapiro et al., 1982).
Different trials were undertaken to determine whether these screening
programmes achieved their goals. The eight most important trials are the
following: Chu et al. (1988), Alexander et al. (1999), Bjurstam et al. (1997), Frisell
et al. (1997), Tabár et al. (1995), Miller et al. (1992a, 1992b), Andersson et al.
(1988), and Andersson & Janzon (1997). Most of these trials show a significant
reduction in breast cancer mortality, especially for women aged between 50 to 70
years. These results have been used to guide screening programmes worldwide.
The efficacy of screening mammography, especially for women in the age group
from 40 to 49 years, remains controversial. Due to this reason, the American
43
Cancer Society (ACS) has recommend mammography screening on an annual basis
for all women beginning at the age 40 (American Cancer Society, 2009).
2.4.1 Errors In Screening
Several studies have shown that approximately 20 percent of all interval
carcinomas are visible on previously screened mammograms (Savage et al., 1994),
(Vitak, 1998), (Burrel et al., 1996). Of all cancers detected during screening, 20
percent are retrospectively considered actionable on previously screened
mammograms (Harvey et al., 1993), (Bird et al., 1992), (Van Dijck et al., 1993).
These numbers suggest that a considerable improvement in mortality reduction is
possible if these errors could be prevented. When mammograms are examined
retrospectively for signs of cancer, the abnormality is considered occult (i.e.,
nothing is visible on the previous mammogram) or classified as a minimal sign or
screening error. An abnormality is called a minimal sign if something abnormal is
found in the Region of Interest (ROI) that is not suspicious enough to recall. If signs
of cancer (malignant tissue) are present that are actionable, it is called a screening
error. However, many tumors do show clear signs of cancer on previously
screened mammograms and many are found by automated detection systems at
high specificity levels. A problem with this type of study is the subjective nature of
the findings: normal, minimal sign and screening error, since the definitions of the
findings vary considerably between radiologists.
There are two reasons as to why women with a visible tumor are not recalled for a
follow up. The first possibility is that the sign was overlooked, and was not
examined at all by the radiologist. The second possibility is that the sign was
examined but it was considered benign, normal (no tumor), or not found
44
suspicious enough for further examination. So far, only a few studies have focused
on the reasons as to why errors are made in the field of mammography
(Hartswood et al., 1998), (Hutt, 1996), although some work has been done in other
medical areas (Friedman, 1999). Much work on the signal detection theory (Green
& Swets, 1966) has been done in psychology departments, some related to the
medical field (Laming, 1995). Psychophysical evidence exists that inserting extra
abnormal signals to increase the target rate improves the performance when the
target rate is very low, which is the case in breast cancer screening programs
(Laming, 1995).
Mammographic signs of breast cancer missed in most screening programs are
mass lesions and architectural distortions (Bird et al., 1992), (Burrel et al., 1996),
(Vitak, 1998). Masses are often obscured by glandular tissues or have low contrast
or no clear cancer (malignant) signs, like fuzzy edges or spicules. MCCs on the
other hand are more easily detected by radiologists, but are often hard to classify
between benign and malignant types.
2.5 Imaging Modalities
At the moment the modality of choice for breast cancer screening is digital
mammography. For further examinations, or when digital mammography is not
sufficient, other modalities are used, which include: ultrasonography (US) and
magnetic resonance imaging (MRI). The following section gives an overview of
different imaging modalities used for mammography screening.
45
2.5.1 Digital Mammography
Although most radiologists are more comfortable with the use of screen-film
combinations, its disadvantages are obvious. Once an image is printed using
screen-film technology, it can no longer be manipulated, and any information
available in digital format, but not captured on the printed image will be lost.
Furthermore screen-film combinations have limitations in detecting subtle soft
tissue lesions, especially during the presence of dense glandular tissue (Kobatake
et al., 1998).
To overcome the limitations of screen-film mammography, digital mammography
was introduced (Lewin et al., 2001). Digital mammography provides several
advantages over screen-film mammography such as the easy access to images, the
use of computer-aided detection methods, improved means of transmission and
retrieval and storage of images, and the use of a lower average dose of radiation
without compromising the diagnostic accuracy.
In a recent study, the authors in Pisano et al. (2005) compared the diagnostic
accuracy of digital and screen-film mammography. In this study a total of 49,528
asymptomatic women underwent both digital and film mammography. Breast
cancer status was ascertained by a breast biopsy or a follow-up mammogram. This
study showed that the overall diagnostic accuracy of digital and film
mammography was similar, digital mammography however turned out to be more
accurate in: women under the age of 50 years and women with radio-graphically
dense breasts. As discussed in Section 1.1 previously, a notable advantage of digital
mammography is the possibility of computerized detection of breast cancer.
46
2.5.2 Ultrasonography
The role of Ultrasonography (US) in breast imaging is a subject of ongoing
discussion. Studies that have been performed using US as a mammogram screening
tool failed to establish it’s the efficiency of US. Thus, it has been concluded that US
should not be used as a mammogram screening tool (Rahbar et al., 1999).
2.5.3 Magnetic Resonance Imaging
High resolution Magnetic Resonance Imaging (MRI) of the breast has recently
emerged as a sensitive instrument for the detection of breast cancer. MRI has
proven to be useful in screening younger women with dense breasts who are at a
high risk of developing breast cancer (Stoutjesdijk et al., 2001). MRI can also be
used as an adjunct to mammography for selected patients. However, MRI has a
significant false positive (FP) rate (see Section 3.3.6) and it is not available in all
areas due to being more expensive than digital mammography. Other limitations of
MRI are that, it requires contrast injection and it can cause problems with
claustrophobia. Thus, at the moment MRI remains limited to specific problem
solving situations for patients at high risk for cancer.
2.6 Mammogram Analysis Using Digital Mammography
Digital mammography is an accepted and often preferred screening modality to
detect breast abnormalities (Powell & Stelling, 1994). An X-ray passes through the
breast, being absorbed selectively by different tissue types and emerges to be
recorded onto a film or plate (Roebuck & Blamey, 1990). Screen-film
mammography is generally the most common form today, where the X-rays strike
a screen which emits photons that expose a photographic film (Dance, 1993). Over
the years the radiation dose to the patient has decreased. Radiation exposure is
47
measured either in Roentgen �� or Coulomb per kilogram (C/kg), where one
= 2.58 ×10/� C per kg. The first dedicated mammography unit introduced
commercially in 1969 typically delivered an 8 to 12 patient dose. By 1976 a
screen-film system was introduced which lowered this dose to approximately
0.08 (Andolina et al., 1992). At the same time as radiation dose decreased, image
quality increased, through anti-scatter grids that absorb scattered X-rays and
increase contrast of the image and provide improved compression systems and
automatic exposure (Powell & Stelling, 1994).
There are alternatives to the screen-film mammography. Xeromammography, now
an obsolete technique, was established when non-screen films were being used
(Roebuck & Blamey, 1990). Just as the screen-film combination displaced
Xeromammography, recent digital detectors have the potential to become the
dominant technology due to the advantage of generating digital images; digital
mammography.
2.6.1 Breast Positioning in Digital Mammography
In digital mammography, the breast is compressed between two parallel plates to
spread the breast tissue and make the breast a block of uniform physical thickness
for the X-rays to pass through. This compression can be performed at different
angles to generate different orientations of the breast. The two standard views
used in screening mammography are the: Cranio-Caudal (CC) view, generating a
top to bottom view of the breast and the Mediolateral Oblique (MLO) view, a side-
on view at approximately 45°. Examples of each view are shown in Figure 2.10
with the CC view shown in Figure 2.10(a) and MLO view in Figure 2.10(b).
48
While the breast is compressed to a uniform physical thickness during
mammography, the radiographic density of each tissue type present in the breast
determines the appearance of the mammogram. Radiographic density is the term
used to describe the level of attenuation that the X-rays experience from the source
to the detector. The higher the density, the less developed the film, resulting in
appearance from fully exposed (black) to unexposed (white) depending on the
tissue type. The fat in breasts has a low density, allowing the X-rays to pass
through easily to expose the film, hence fatty areas of the mammogram are dark, in
some places almost as dark as portions of the image where there is no tissue
(background pixels). The glandular tissue in the breast has a somewhat higher
density, resulting in brighter areas, as does the tissue of the pectoral muscle.
Microcalcifications are very high in density; some lesions also have high density
characteristics.
(a) (b)
Figure 2.10: The two standard views of the breast used in screening
mammography (a) The craniocaudal view, from the head down (CC view)
(b) A mediolateral oblique view, with the breast viewed from the side (MLO view)
49
As X-rays pass through a 3D breast to create a 2D image, the brightness of a pixel
on the mammogram often represents the superposition of a number of tissue types
that the X-ray has passed through on its way. Superposition presents several
problems for segmentation of the breast tissue. It also presents problems with
diagnosis. Microcalcifications are of much higher density than a comparable
volume of glandular tissue. However, when superimposed on a large amount of
glandular tissue, the ‘bright’ microcalcifications also appear less significant
(Roebuck & Blamey, 1990).
2.6.2 Breast Regions in Digital Mammograms
Regions present in the mediolateral oblique (MLO) view of a mammogram (see
Figure 2.10(b)) are now identified to give the reader a chance to understand what
part of the image is being discussed when certain names are used. The simplest
distinction is to differentiate the image between breast and non-breast regions.
The non-breast regions include the image background, labels, scanning artifacts
and tapes, which might be superimposed over the breast region.
(a) (b)
Figure 2.11: Mammogram decomposition (a) Original mammogram image
(b) Attempted decomposition of mammogram in (a) into separate breast regions
50
(a) (b)
Figure 2.12: Magnified views showing breast nipple (a) Nipple in the breast
profile (b) Breast profile without the nipple
While non-breast regions are straight forward to identify, the division between
regions inside the breast is less distinct due to the overlap of tissue types. An
attempt is made in Figure 2.11 to indicate different breast tissues. The
mammogram can be divided into fatty tissue (dark and glandular tissue), which is
high in pixel intensity. The pectoral muscle is a characteristic feature of the MLO
view of a mammogram and presents itself as a bright triangle in the top left or
right corner of the mammogram (Kwok et al., 2004).
Another feature is the nipple which may not necessarily be seen in the breast
profile in Figure 2.11. A close-up image with the nipple in the breast profile and
breast profile without a nipple is shown in Figure 2.12. Near the border of the
breast lies a region of fatty tissue termed as the near-skin tissue. This region
results due to the poor compression of the breast at the edge causing a gradual
decrease in thickness towards the skin-line. The near-skin tissue region is not
usually visible in an original mammogram image without contrast enhancement.
Figure 2.13 shows an original mammogram image and an enhanced version
51
(mask) of the original image indicating the near-skin tissue boundary of the breast
profile (Wirth et al., 2007).
(a) (b)
Figure 2.13: Near-skin tissue of a mammogram barely visible in mammogram
(a) Original mammogram image
(b) Enhanced image indicating the near-skin tissue as the mask of (a)
2.6.3 Types of Breast Tissues
The appearance of the breast tissue in a mammogram varies between images. The
process of involution leads to the change from a predominantly bright, glandular
tissue filled image in younger women to a darker, mostly fatty image (Roebuck &
Blamey, 1990). Breasts can be divided into a number of classes based on the
appearance of the glandular tissue (Wolfe, 1976). Breast tissues are classified into
three major types based on their density: (i) fatty, (ii) fatty-glandular, and (iii)
dense-glandular. An example of the different densities of glandular breast tissues is
shown in Figure 2.14.
52
Figure 2.14: Three mammograms images with different breast tissue densities.
From left to right: fatty, fatty-glandular and dense-glandular.
2.6.4 Present Clinical Protocol
This section presents the clinical protocol followed by radiologists for
mammographic examination and interpretation. Standardized mammographic
interpretations follow from the Breast Imaging Reporting and Data System (BI-
RADS) lexicon. The American College of Radiology (ACR) developed the BI-RADS as
a measure for mammographic interpretation for radiologists. The BI-RADS
provides a mechanism for describing the characteristics of a given abnormality
including the final pre-pathology finding (BI-RADS, 2010).
For classification of mass lesions the borders, shape and relative intensities of the
lesions are important descriptive features. In the following subsections, the
relevant BI-RADS descriptors and assessment categories are presented.
53
2.6.4.1 BI-RADS Descriptors and Assessment
BI-RADS descriptors are important factors for predicting malignancies that are
assessed and provided by radiologists. Mass narratives include the overall shape
description, the border region margin regularity and the relative intensity of the
mass region compared with the ambient normal tissue intensity. The BI-RADS
lexicon provides a four-category rating for assessing the overall breast tissue
characteristics in terms of fibro-glandular composition (BI-RADS, 2010). The
composition categories relate to the degree of interpretation difficulty. Similarly,
the BI-RADS gives a five-point overall assessment that is related to the degree of
probable malignancy or necessary follow up work.
(a) (b)
Figure 2.15: BI-RADS mass descriptors for (a) shape (b) margin
2.6.4.1.1 BI-RADS Mass Descriptors
BI-RADS have established mass descriptors such as for shape and margin, for the
detection of mass lesions as indicated in Figure 2.15. The shape and margin
properties of the BI-RADS mass descriptors are as follows (BI-RADS, 2010):
54
• Shape: The shape of the mass is described with a five-point assessment:
round, oval, lobular, irregular and architecturally distorted as shown in
Figure 2.15(a).
• Margin: The mass margins modify the boundaries. For example the overall
shape of the mass may be round, but close inspection may reveal scalloping
along the border, which may indicate a degree of irregularity or a lobular
characteristic. The margins are rated with a five-point system:
circumscribed (well-defined/sharply-defined) margins, obscured margins,
micro-lobulated margins, ill-defined margins and spiculated margins as
shown in Figure 2.15(b).
2.6.4.1.2 BI-RADS Assessment Categories
The BI-RADS assessment categories are defined for standardized interpretations of
mammographic findings. Each category provides the overall assessment related to
the findings and the necessary follow up required. The assessment categories are
summarized in Table 2.1 (BI-RADS, 2010).
2.6.5 Mammogram Interpretation
Radiologists interpret mammographic examination in the form of a mammogram
report. A mammogram report describes the findings i.e. breast abnormalities, and
provides the radiologist’s impression based on BI-RADS with recommendations on
the appropriate actions to be taken. The elements of the mammogram reporting
from BI-RADS are shown in Table 2.2 (BI-RADS, 2010).
55
Table 2.1: BI-RADS assessment categories (BI-RADS, 2010)
Category Assessment and
Recommendation Findings
Category 0 Incomplete Assessment (No recommendation)
Needs additional imaging evaluation.
Category 1 Negative
(No action required) The breasts are symmetrical and no abnormalities are present.
Category 2 Benign Finding
(No action required) This is a negative mammogram and no abnormalities are present.
Category 3 Probably Benign Finding (Short interval follow-up
suggested)
A finding has a high probability of being benign.
Category 4 Suspicious Abnormalities
(Biopsy should be considered)
These are lesions that do not have the characteristic morphologies of breast cancer but have a definite probability of being malignant.
Category 5 High Suggestive of Malignancy (Appropriate action should be
taken)
These lesions have a high probability of being malignant.
2.7 Summary
This chapter discussed breast cancer detection and digital mammography. General
information about the structure and functions of the breast, breast tumors and
literature regarding breast cancer screening were presented at first. Next,
background issues concerning different imaging modalities such as digital
mammography, ultrasonography (US) and magnetic resonance imaging (MRI)
were reviewed and discussed. Towards the end of this chapter, the analysis and
interpretation of digital mammograms using the BI-RADS lexicon was discussed.
56
Table 2.2: Elements of mammogram reporting from BI-RADS
Mammogram
Element Description
Findings
Breast abnormalities (i.e. mass lesions and MCCs) found from mammograms in terms of size, location and shape characteristics.
• Primary signs of breast cancer: spiculated masses and clustered pleomorphic MCCs.
• Secondary signs of breast cancer: asymmetrical tissue density, skin thickening, retraction and focal distortion of tissue.
Impression Contains the radiologist’s overall assessments (findings and breast abnormalities) using the BI-RADS lexicon.
Recommendation
Depending on the assessment, the recommendation contains specific instructions on what actions should be taken. For example, in Table 2.2, the radiologist could recommend:
• For Category 0: Additional imaging such as spot views, MRI etc.
• For Category 1 and 2: No action is necessary. • For Category 3: A six months follow-up procedure
is required to establish the finding’s stability. • For Category 4 and 5: A biopsy is required.
57
CHAPTER 3
ELEMENTS OF COMPUTER-AIDED DETECTION
3.0 Overview
This chapter presents and discusses the background literature on the computer
processing of digital mammograms. In Section 3.1, computer-aided detection
systems are introduced with the literature review of computerized breast cancer
detection techniques presented in Section 3.2. Section 3.3 highlights and identifies
the key techniques and algorithms used in this research to develop a framework
for the computerized detection of breast cancer. Section 3.4 presents the
fundamental concepts of digital image processing with emphasis on image
segmentation techniques used in digital mammography applications. Lastly,
Section 3.5 emphasizes on the use of texture-based analysis for the purpose of
feature extraction in pattern classification problems.
3.1 Computer-Aided Detection Systems
In recent years major effort has been made to develop digital mammography
applications which can assist radiologists in the detection and characterization of
malignant and benign abnormalities. One such application of digital
mammography is computer-aided detection, as discussed in Section 1.1 previously.
Computer-aided systems identify and mark suspicious regions on mammograms to
bring them to the attention of radiologists. These systems minimize search,
perception and interpretation errors even if radiologists fail to recognize
58
suspicious abnormalities. Computer-aided detection is intended to be used after
the radiologist has completed evaluation of the mammographic images and has
made an initial decision whether patient recall is required (Hutt, 1996). As an
example, if the radiologist identifies an abnormal region on a mammogram during
initial reading and that area does not get marked by the computer-aided method,
the radiologist is advised to interpret the mammogram as positive and to recall the
patient for further work up. Since, computer-aided detection is proposed as an
adjunct to digital mammography to decrease search, detection and interpretation
errors (see Section 1.1), the radiologist makes the final decision if a clinically
significant abnormality exists and decides whether further diagnostic evaluation is
warranted (Jirari, 2005).
The hope lies in the fact that computer-aided detection systems will improve the
sensitivity of digital mammography without substantially increasing patient recall
rates (Bozek et al., 2008). The following section provides the review of
computerized detection of breast cancer (mass lesions and MCCs) in digital
mammography applications.
3.2. Review of Computerized Breast Cancer Detection Techniques
As it is known, the goal of computerized breast cancer detection in digital
mammography is to identify the presence of abnormalities such as mass lesions
and MCCs. Lesion detection is possible from a single mammogram image, as is the
detection of MCCs, which is subjected to a wide volume of publications. Whichever
way they are detected, masses and MCCs need to be classified into their malignant
and benign types; a technique many authors have tried to automate.
59
Many papers have been published on enhancing digital mammograms for optimal
viewing, mainly for computer-aided detection of MCCs (Mutihac et al., 1998),
(Strickland et al., 1996) and (Cernadas et al., 1996) and masses (Woods & Bowyer,
1996). Most of these studies provide evidence that the radiologists perform better
on computer enhanced images. Important work was done by Aylward et al. (1998)
and Netsch et al. (1998) to transform mammograms in such a way that they can be
printed or examined on a monitor optimally. For example, the dark area near the
skin line can be enhanced (Karssemeijer & te Brake, 1996), (Bynd et al., 1997) and
the pectoral muscle can be filtered out (Nicolaou et al., 2008), largely reducing the
intensity range in the mammogram. Good contrast will be available in the whole
Region of Interest (ROI), both in the pectoral area as well as near the skin line.
Furthermore, the authors in Highnam et al. (1996) described a method to filter
scatter from digital mammograms.
Both mass and MCC processing follows a similar set of mammogram preprocessing
steps. At first, the mammogram images are enhanced to highlight some property of
the desired regions (ROIs) through either spatial filtering or time-frequency based
methods (Thangavel and Karnan, 2005b). Heuristic features are then computed
from the enhanced images (ROIs) and some basic classification is performed to
differentiate between masses (Woods & Bowyer, 1996) and MCCs (Bazzani et al.,
2001), (Lu & Bottena, 2001). Further classification can be performed to identify
masses and MCCs into malignant and benign types (Veldkamp et al., 2000),
(Rahbar et al., 1999), (Jiang et al., 1998), typically using machine learning
frameworks such as ANNs, statistical classifiers such as SVMs, or some sort of
decision tree mechanism. During classification, features such as: shape, size, and
60
texture properties (such as statistical distribution of regions and other measures
of texture) of the ROIs should be taken in account (Martì et al., 2003).
The search for abnormalities (masses and MCCs) in mammogram images generally
uses the breast profile boundary to constrain processing to the breast area only,
avoiding spending time processing non-breast regions. In order to constraint
mammogram processing to the breast tissue area only, mammogram segmentation
is a fundamental step to suppress un-important regions in mammograms (Wirth et
al., 2007). Knowledge of the pectoral muscle may also be used in single image
analysis, referred to as bilateral comparison. As breast tumors are frequently
projected in the lower areas of the pectoral region and the presence of the
intensity gradient at the pectoral muscle edge may easily generate false alarms or
miss tumors (Karssemeijer & te Brake, 1998), thus the location of the pectoral
muscle may be used to subtract its contributing intensity from the image.
Many attempts have been made to identify mass lesions and MCCs (Cascio et al.,
2008), (Domínguez & Nandi, 2008), (Li et al., 1997), (Song et al., 2009), (Wirth et
al., 2007), (Martì et al., 2003) in order to classify between malignant and benign
types. Approaches to mass and MCC detection have been based on concepts which
mainly include: template matching (Özekes et al., 2005), wavelets (Soltanian-
Zadeh et al., 2004), (Mousa et al., 2005), (Gorgel et al., 2009), and measures of
texture (Varela et al., 2001), (Oliver et al., 2007), (Mudigonda et al., 2001), (Martì
et al., 2003), (Lyra et al., 2008), (Karahaliou et al., 2008), (Bovis & Singh, 2002).
The list of methods used is extensive, only recent approaches for detection of mass
lesions and microcalcifications/MCCs are discussed in the following section.
61
3.2.1 Detection of Microcalcifications/MCCs
The detection of microcalcifications is an important topic in computerized
detection of breast cancer, because it is a task that radiologists also find
challenging. To achieve a good positive predictive value (PPV) it is important to be
able to discriminate between malignant and benign abnormalities, because only 20
percent of all microcalcification clusters (MCCs) are due to malignant processes.
Pointing out all microcalcifications is a tedious and time consuming task, not suited
for use in clinical practice. So far only few completely automated methods have
been published (Sorantin et al., 1998).
The authors in Kaufmann et al. (2001) defined a MCC as three or more
microcalcifications within a 1cm diameter circle with cluster features such as:
calcification number, cluster area, and statistical measures of inter classification
distance combined with individual calcification features. These features were used
to classify MCCs into benign and malignant categories by evaluating the k-nearest
neighbor and Bayesian classifiers (Kaufmann et al., 2001). Similar sets of MCCs
features have been used by Jiang et al. (1998). The authors in Jiang et al. (1996)
developed a method that outperformed five radiologists, using an ANN that
classified MCCs based on eight features that were computed for each cluster.
Bottema and Slavotinek (2001) determined the convex hull of a MCC by using the
ratio of the maximum and minimum distances from the convex hull boundary to its
geometric centre in order to separate clusters of large DCIS calcification from
benign calcifications.
Microcalcifications in digital mammograms represent a sharp transition in
intensity as they are quite bright; hence enhancing mammograms to highlight
62
small objects representing high frequency is used by most approaches (Soltanian-
Zadeh et al., 2004), (Mousa et al., 2005), (Gorgel et al., 2009). The use of wavelets is
a popular method to extract high frequency regions from mammogram images in
order to search for microcalcifications and MCCs (Wang & Karayiannis, 1998),
(Brown et al., 1998), (Bazzani et al., 2001), (Lu & Bottena, 2001). Other techniques
include: Markov random field models (Karssemeijer, 1992) (Veldkamp &
Karssemeijer, 1998), Spatial filters (Diahi et al., 1998), (Gürcan et al., 1998), Box-
rim filters (Bottema & Slavotinek, 1998) and background subtraction with a model
of the background produced by methods as Gaussian blurring (Lu & Bottena, 2001)
or polynomial modeling (Bottema & Slavotinek, 1998), (Lu & Bottena, 2001).
Dealing with noise in mammograms is important for MCC detection algorithms.
Most methods use local adaptive thresholding, because noise levels vary across
mammogram images. Nishikawa et al. (1994) used an initial global threshold level,
followed by a locally adaptive threshold step. The authors in Chitre et al. (1994)
computed a local threshold image and used the local deviation of grey levels as a
threshold to decide whether or not a pixel belonged to a MCC. Karssemeijer (1992)
developed a method for this purpose which was improved by Veldkamp &
Karssemeijer (1998). The advantage of a global correction approach is that the
statistics are much better than for local estimations of the appropriate threshold
level. Several other research groups such as Mutihac et al. (1998) and Strickland &
Hahn (1996a) focused on noise equalization.
The properties of individual microcalcifications are important for their grouping. A
large percentage of mammograms contain benign calcifications (Bottema et al.,
2001). However, the more calcifications there are per unit area, the more likely is
63
the abnormality to be malignant (Roebuck & Blamey, 1990). Automated
segmentation of microcalcifications can lead to complicated algorithms using more
than one of the methods outlined above. As an example, the authors in Doi et al.
(1993) and Yoshida et al. (1994) followed a series of steps which included: wavelet
transform based processing, global thresholding, morphological erosion, local
thresholding, texture analysis and clustering. As in the above example, using a
variety of operations leads to a number of parameters leading to the problem of
tuning due to interdependence of the parameters. However, adaptive tuning of
parameters can be used to improve the success rate of the algorithm while keeping
the False Positives (FPs) less. During breast cancer screening, the incorrect
identification of a malignant abnormality as benign in breast cancer patients
generally is referred to as a false positive (FP) (Kocur et al., 1996), (Fogel et al.,
1998) (see Section 3.3.6). Similar to breast cancer screening, the performance of a
medical diagnostic test in digital mammography applications is typically measured
using the Receiver Operating Characteristic (ROC) curve analysis as presented in
Section 3.3.6, where the four performance measures: true positive (TP), false
positive (FP), true negative (TN) and false negative (FN) measure the sensitivity
and specificity of the tested samples.
The authors in Anastasio et al., (1998) used Genetic Algorithm (GA) to tune the
parameters in the case of the algorithm of Doi et al. (1993), which led to a
sensitivity increase from 80 to 87 percent with a FP rate of 1 per image. Kobatake
et al. (1998) used subtraction with several smoothed images, each produced with a
top-hat transformation (using morphological operations) using several structuring
elements (STRELs) to increase sensitivity and to remove the False Positives (FPs)
causing ‘elongated shadows’ due to glands and blood vessels in the breast tissue.
64
The authors in Highnam and Brady (1999) proposed another way to remove FPs,
is by performing direct segmentation of the curve-linear structures in a
mammogram. Other methods to segment potential microcalcifications and MCCs
include: mathematical morphology (Zhao et al., 1992), (Dengler et al., 1993),
(Hagihara et al., 2001), (Kaufmann et al., 2001), (Bruynooghe, 2001), directional
recursive median filtering (Cernadas et al., 1998), Fuzzy logic (Cheng et al., 1998),
and fractal theory (Lee et al., 2000), (Li et al., 1997).
Recent approaches for identification of MCCs typically use features such as: shape,
size and texture based properties (Sorantin et al., 1998), (Meersman et al., 1998),
(Lu & Bottema, 2001), (Brown et al., 1998), (Martí et al., 2001), (Jiang et al., 1998)
applicable for pattern classification. The authors in Lee et al. (2001) compared four
methods of microcalcification detection: (i) Karssemeijer’s Markov random field
(Karssemeijer, 1992), (ii) Strickland’s wavelet based algorithm (Strickland, 1996),
(iii) Grey-level “isophote” contours (Guillemet et al., 1996), and, (iv) Adaptive
threshold based method (Wallet et al., 1997). The algorithms were tested on
images from two databases, with better results for one database. The criterion for
success was detecting at least 80 percent of the MCCs with as little FPs as possible.
The algorithms by Wallet et al. (1997) and Guillemet et al. (1996) did not meet the
80 percent sensitivity requirement. In conclusion to the experiment performed by
Lee et al. (2001), Karssemeijer's algorithm (Karssemeijer, 1992) is favored more as
it produces less FPs than Strickland’s wavelet based algorithm (Strickland, 1996).
3.2.2 Detection of Mass Lesions
A large variety of techniques have been applied to the problem of mass lesion
detection, but most follow a two-step scheme that was described by Woods and
65
Bowyer (1996). First, one or more features are computed for each pixel, after
which each pixel is classified and the suspicious pixels are grouped into a number
of suspicious regions. In the second step, these regions are classified as normal or
abnormal regions, based on regional features like size, shape, contrast and texture
properties (Woods and Bowyer, 1996). Two signs can indicate the presence of a
lesion: (i) a radiating pattern of spicules, or (ii) a central mass. To detect the whole
range from architectural distortions to circumscribed mass lesions, both signs
must be detected (Zhang & Giger, 1995), (Cascio et al., 2008).
3.2.2.1 Central Mass
The central mass of a lesion is a circular bright region with a diameter between
5mm and 5cm. Convolution of mammogram images with a zero-mean filter with a
positive center and a negative surrounding area was used by a number of research
groups to detect mass lesions (Sahiner et al., 1996), (Zheng et al., 1995), using the
Laplacian of the Gaussian (LoG) and Difference of Gaussian (DoG) filters. This is an
easy and intuitive approach to detect bright blobs, but may not be suited to find
masses with lower contrast. Other approaches that are less dependent of the
contrast are more useful, like template matching, a method used in some early
research studies for the detection of the central mass (Ng & Bischof, 1992), (Lai et
al., 1989). In template matching, a model is made of the appearance of a mass and
the mammogram is searched for regions that resemble this model. This approach
is more related to the shape of the region, rather than the contrast. It is especially
hard to detect low contrast masses using this method, however it may outperform
convolution based approaches.
66
Other approaches for lesion detection focus on the analysis of the gradient
patterns in the region of interest (ROI). The appearance of masses in
mammograms varies and therefore the rigid approach of template matching is not
very successful. In an area with a central mass, the orientation of the gradients will
be towards the center of the mass. Statistical analysis of this type of pattern can be
used to discriminate masses from other structures. The authors in Groshong and
Kegelmeyer (1996) used a generalized Hough transform for detection of the
central mass. The strongest edges in the ROI are accumulated in a Hough space
where each location relates to a center and a radius. Masses yield peaks in this
space (Groshong & Kegelmeyer, 1996). The authors in Zwiggelaar et al. (1999)
applied a one-dimensional recursive median filter over a number of different
angles to each pixel in order to detect the central mass.
Typically masses look very much similar to normal glandular structures, and they
are only detectable due to asymmetry between the left and right breasts. Matching
two breasts is a complicated procedure because there is only an approximate
correspondence between the normal tissue in the left and right breasts and due to
variations in compression and positioning, the variation in appearance is even
made larger. Yin et al. (1991, 1993) applied a simple rigid body transform to align
the skin line of the two (left and right) breasts. Another sophisticated approach
includes matching corresponding points between the two breasts. Lau et al. (1991)
use a set of 3 control points and an estimation of the nipple. Sallam and Bowyer
(1994) used a more general warping method to match automatically detected
landmarks of the glandular tissue. When the two breasts are correctly matched,
subtraction and smoothing can be done to find a number of suspicious regions
(Lau & Bischof, 1991), (Yin et al., 1991), (Yin et al., 1993).
67
More recent approaches towards detection of lesions focuses on texture based
analysis (Varela et al., 2001), (Soltanian-Zadeh et al., 2003). Haralick’s texture
descriptors (Haralick et al., 1973), (Haralick, 1979) and Laws’ texture measures
(Laws, 1980) have been used quite extensively to detect mass lesions. Kegelmeyer
Jr. (1994) used Law’s texture measures together with the Analysis of Local
Oriented Edges (ALOE) as features to determine the probability of abnormality for
each pixel using a binary decision tree. Polakowski et al. (1997) also made use of
Law’s texture measures using a system based on several modules enhancing
masses through a Difference of Gaussian (DoG) filter and calculating Law’s texture
measure along with features based on shape, size, and contrast with an ANN for
classification of regions as malignant or benign. In more recent times, Khuzi et al.
(2009) applied Haralick’s texture descriptors for the identification of masses in
digital mammograms. It has been reported that the success of lesion detection
using Haralick’s texture descriptors is highly dependent upon mammogram
preprocessing (Khuzi et al., 2009)
3.2.2.2 Spicules
Architectural distortions are a straight forward indication of a malignant lesion.
When a mass is surrounded by spicules, it is likely to be malignant. Many stellate
lesions are easier to detect by their spicules than by their central mass.
Kegelmeyer et al. (1994) computed histograms of local gradient orientations,
which indicated that areas with a spicule pattern have flatter histograms than
normal areas. This feature was combined with four texture features; however, the
good results of Kegelmeyer et al. (1994) could not be reproduced by other groups
(Woods & Bowyer, 1996). Analysis of texture in the Hough space was the basis of
an approach proposed by (Zhang and Giger, 1995). The authors in Parr et al.
68
(1996a, 1996b) developed a model for detection of spicules using the Principle
Component Axis (PCA) and achieved good results for the detection of spicules.
Other techniques include pixel-level algorithms that detect both spicules and
masses, but are very sensitive and are known to signal many false positives (FPs)
(see Section 3.3.6).
3.2.2.3 Normal and Abnormal Regions
The classification of normal (malignant) and abnormal (benign) regions is a well-
studied subject in digital mammography (Veldkamp et al., 2000), (Rahbar et al.,
1999), (Jiang et al., 1998). Articles reviewed on this topic focus on edge analysis of
masses, where vague or spiculated edges indicate malignancy, and sharp and well
defined contours are likely to belong to a benign abnormality (Martì et al., 2003).
The authors in Rangayyan et al. (1997) indicated that radiologists outlining lesions
may induce a bias because the way they outline spicules or vague regions can be
incorporated in the computed lesion features. Other interesting work was done by
Giger et al. (1994) and Hui et al. (1995) who used a radial edge gradient method to
discriminate malignant and benign lesions. Pohlman et al. (1996) developed a
feature describing the tumor boundary roughness for differentiation between
benign and malignant masses.
A number of articles have been published on the topic of discriminating real
lesions from suspiciously looking normal tissue (Sahiner et al., 1996), (Wei et al.,
1995), (Polakowski et al., 1997), (Yin et al., 1993). In most cases, features or
heuristics are computed over a large region containing the suspicious area (Wei et
al., 1995), however, most research groups have segmented the suspicious area to
reduce the number of FPs. (Domínguez and Nandi, 2009), (Martì et al., 2003).
69
Segmentation of the suspicious breast tissue is useful in separating the abnormal
breast tissue from the breast normal tissue, as it enables computation of features
related to the edge of the region, as well as other properties such as the contrast,
shape and size. The segmented suspicious area is typically known as the Region of
Interest (ROI). Recent approaches by authors in (Khuzi et al., 2009) and (Veldkamp
et al., 2000) signify that using texture based features (Haralick et al., 1973),
(Haralick, 1979), (Laws, 1980) a considerable improvement in classification
between malignant and benign types is generally achieved by the removal of FP
signals.
3.3 Computerized Detection of Breast Cancer
The goal of computerized breast cancer detection systems is to reduce the number
of false positive (FPs) and to achieve high sensitivity for detecting cancers that
radiologists might miss (Hutt, 1996). During breast cancer screening, the incorrect
identification of malignant lesion as benign in breast cancer patients, generally
referred to as a false positive (FP) (Kocur et al., 1996), (Fogel et al., 1998) as
discussed in Section 3.3.6. The clinical utility of computerized systems depends
upon the number of FPs per image, since radiologists must take extra time and
care to inspect areas of the mammograms with FPs (Nagel et al., 1998).
Computer processing of mammograms in digital mammography applications
typically involves image processing. Similarly, for pattern classification between
malignant and benign abnormalities in mammograms, classification (machine
learning) is required. Figure 3.1 shows a typical framework of a computerized
breast cancer detection system for digital mammography applications. The
70
following sections discuss in detail each of the six stages in the computerized
detection of breast cancer (Hutt, 1996).
Figure 3.1: General framework the computerized detection of breast cancer
(Hutt, 1996)
3.3.1 Image Preprocessing
Preprocessing of digital mammograms generally involves noise and radiopaque
artifact suppression and image adjustment. Image adjustment is usually achieved
by performing contrast enhancement using image histogram measures. Increasing
the contrast is very essential in mammograms, especially for dense breasts (Wang
& Karayiannis, 1998). Contrast between the malignant and benign abnormalities
maybe present in mammograms but may not be discernable to the human eye. As a
result, differentiating between malignant and benign abnormalities is difficult
(Wang & Karayiannis, 1998). Conventional image processing techniques may not
work well on mammography images because of the large variation of feature sizes
and shapes (Morrow et al., 1992). There are a few approaches to enhancing
mammographic features, as reviewed by the authors in Thangavel et al. (2005a).
Machine Learning
Image Processing
Image Segmentation (Stage 2)
Image Preprocessing
(Stage 1)
Feature Extraction (Stage 3)
Feature Selection (Stage 4)
Classification (Stage 5)
Performance Evaluation (Stage 6)
Digital
Mammograms
Result
71
One technique is to suppress background and digitization noises and the other is to
increase the contrast of suspicious areas (Thangavel & Karnan, 2005b).
Noises due to intrinsic characteristics of an imaging device or imaging process will
impact the sensitivity of the classification result. Several types of filters have been
reported by Qian et al. (1994) in order to reduce imaging and digitization noises.
Methods like straight line windowing (Chan et al., 1987) and hexagonal windows
(Glatt et al., 1992) have been introduced as non-linear filtering techniques. Though
non-linear filtering techniques have shown to be more successful for noise
suppression than linear approaches, they do not necessarily show significant
improvements in image detail preservation. Other techniques include Median
filtering and edge-preserving. Median filtering locally adapts to grayscale images
using an empirically derived threshold criteria (Lai et al., 1989) and (Thangavel &
Karnan, 2005b). Other techniques include edge-preserving smoothing, which
searches for a homogeneous neighborhood in different directions of a given pixel
and averages this neighborhood.
Mammographic artifacts are small emulsion continuity faults on X-ray
mammogram films, which look like MCCs. Artifacts in mammograms normally
contain the form of labels, markers and wedges in the unexposed air-background
(non-breast) region. Such artifacts are usually radiopaque in the sense that they
are not transparent to radiation. These artifacts are usually sharply defined and
brighter than the microcalcifications, which are normally present in the
background region of mammograms. One of the problems with the precise
segmentation of the breast profile in mammograms is due to the existence of
artifacts (e.g. a label overlapping the breast region), which often results in a non-
72
uniform background region causing many segmentation algorithms to fail. A
robust artifact suppression algorithm based on area morphology developed by
Wirth et al. (2004) removes radiopaque artifacts from the background region of
mammograms (Yapa & Harada, 2008), (Wirth et al., 2007). Thus, the approach
proposed in this research applies the 2D median filtering (Thangavel & Karnan,
2005b) for the purpose of noise removal and the algorithm based on area
morphology (Wirth et al., 2004) is used for the purpose of radiopaque artifact
suppression.
(a) (b)
Figure 3.2: Digital mammogram (a) Original mammogram (b) Segmentation of the
mammogram into the breast area (grey) including the pectoral muscle (white) and
background region (black)
3.3.2 Image Segmentation
Following noise suppression and artifact suppression, image segmentation is used
to identify suspicious areas (ROIs) in the mammogram images. The aim of
segmentation is to obtain ROIs containing all breast abnormalities from the breast
tissue and locate suspicious lesions and MCCs from the ROIs. Mammographic
abnormalities such as mass lesions are extremely difficult to identify because their
73
radiographic and morphological characteristics resemble those of normal breast
tissues. Since a digital mammogram is a projection image, mass lesions do not
appear as isolated densities but as overlaid over parenchymal tissue patterns.
During the last decade a number of mammogram segmentation techniques have
emerged. These segmentation techniques include: Region growing (Kupinski &
Giger, 1998), (Petrick et al., 1999), (te Brake & Karssemeijer, 2001), Markov
Random Fields (Li et al., 1995), Fractal modeling (Lefebvre et al., 1995), (Li et al.,
1997), Tree structured wavelet transform (Qian et al., 1995), Adaptive density-
weighted contrast enhancement (Petrick et al., 1996), Fuzzy logic (Gavrielides et
al., 2000), Morphological operations (Li et al., 2001) and Dynamic programming
based techniques (Domínguez et al., 2007).
The presence of the pectoral muscle in mammogram images effects results of
intensity based image processing methods and can bias procedures in detection of
malignant and benign abnormalities (Kwok et al., 2004), (Raba et al., 2005),
(Nicolaou et al., 2008), (Xu et al., 2007), (Ferrari et al., 2004), (Mirzaalian et al.,
2007), (Bajger et al., 2005). Thus, during mammogram analysis the pectoral
muscle should be suppressed. Figure 3.2 indicates the segmentation of the pectoral
muscle (white) from the breast region (grey). In Section 3.4.3 a review on
mammogram segmentation techniques widely used in this field of research is
provided.
3.3.3 Feature Extraction
During feature extraction, heuristics are computed (calculated) from the
characteristics of the segmented ROIs. The number of features (heuristics) selected
74
for breast cancer detection reported in literature varies with the approach
employed. Features in different image domains such as: morphological, spatial,
texture can be extracted from digital mammograms (Thangavel et al., 2005a).
During feature extraction, the most important characteristics of the ROIs are
studied and analyzed. Important characteristics of ROIs for classification of
malignant and benign tumors (mass lesions and calcifications) as reported by
radiologists are as follows (Veldkamp et al., 2000):
(a) Polymorphism vs. Monomorphism: Calcifications that are malignant tend to
polymorph while benign clusters are mostly characterized by
monomorphous calcifications of uniform size (Lanyi, 1988).
(b) Size and Contrast: Benign calcifications have larger size and contrast
compared to malignant calcifications.
(c) Branching vs. Round and Oval Type: Linear calcifications can be an
indication of DCIS, since such calcifications are located in the glandular
ducts. Benign calcifications are mostly round or oval in shape and are often
located in the lobules.
(d) Orientation: Malignant calcifications often have shapes that are oriented to
the nipple (Lanyi, 1988).
(e) Number: A MCC with very few calcifications is regarded as less suspicious.
Five or more calcifications, measuring less than 1mm in a volume of one
cubic centimeter, are considered to form a MCC (Popli, 2001).
75
(f) Location: About 48 percent of all cancerous processes are located in the
outer upper quadrant of the breast. Lesions found in this quadrant are
more suspicious (Harris et al., 1996).
During the last decade a number of mammogram feature extraction techniques
have emerged and were applied to digital mammography for the detection of mass
lesions and MCCs. These feature extraction techniques include: texture features
(Sahiner et al., 1998), (Mudigonda et al., 2000), (Hadjiiski et al., 2001), radial edge-
gradient analysis (Huo et al., 1995), gray-level image structure features (Dhawan
et al., 1996), morphological-based features (Chan et al., 1998), Wavelet analysis
(Qian et al., 1999), boundary characteristics of tumors (Kobatake et al., 1999),
Fuzzy-neural modeling (Verma & Zakos, 2001) and region registration using
temporal features (Timp & Karssemeijer, 2006).
The reviewed literature indicates that texture-based features (Haralick et al.,
1973) have been most commonly used for the identification of mass lesions and
MCCs. Texture based descriptors proposed by Haralick et al. (1973) and later
implemented by Kramer & Aghdasi (1999), Soltanian-Zadeh et al. (2004), Chan et
al. (1998) and Khuzi et al. (2009) have been indicated to increase the performance
of machine learning algorithms (Makinacı et al., 2005), (Jirari et al., 2005) and
(Mudigonda et al., 2000) such as SVMs and ANNs. Thus, the approach proposed in
this research uses texture-based features for the machine learning modelling.
Section 3.5 discusses in detail the approach applied in texture feature analysis.
76
3.3.4 Feature Selection
For the purpose of pattern classification, it is desirable to use an optimal number of
features for machine learning modeling. Since a large number of features increases
the computational needs, it becomes more challenging to define accurate decision
boundaries in a large dimensional space. This indicates that an optimal subset of
features needs to be selected for the purpose of machine learning.
Feature selection is an important part of any machine learning task. The success of
a classification scheme mainly depends on the features selected and the
information they provide for their role in the model. Some of the features extracted
from the ROIs in the mammographic images are not significant when observed
alone, but in combination with other features they can be significant for
classification. In general, the reason for performing feature selection is three-fold
namely: (a) improving the classification performance of the system, (b) providing
faster and more cost effective classification and (c) providing a better
understanding of the processes that generated the data (Guyon & Eliseeff, 2003).
During the last decade a number of mammogram feature selection techniques have
emerged and have been applied to digital mammography for the detection of mass
lesions and microcalcifications. These feature selection techniques include: Genetic
Algorithm (GA) (Sahiner et al., 1996), (Zheng et al., 1999), Linear Discriminant
Analysis (LDA) and linear regression analysis (Huo et al., 2000), Genetic
programming (Nandi et al., 2006) and Neural-genetic modeling (Verma & Zhang,
2007).
77
There are many notable benefits of heuristic and feature selection. Firstly, feature
selection facilitates data visualization and understanding and reduces the storage
requirements by reducing the training time and improving classification
performance. The discrimination power of the features can be analyzed through
this process. The goal is to eliminate a feature if it gives little or no additional
information, beyond that subsumed by the remaining features (Koller & Sahami,
1996). In addition, features with high correlation can be eliminated during this
process in order to reduce the overall processing time without affecting the
accuracy of the classifier. Only a few features may be useful or ‘optimal’ while most
may contain irrelevant or redundant information that may result in the
degradation of the classifier’s performance. Irrelevant and correlated feature
attributes are detrimental because they might contribute noise and can interact
counter-productively to a classifier induction algorithm (Hsu et al., 2002).
There are several feature selection techniques that have been well researched and
published. All these methods determine the relevancy of the generated features
towards the classification task. There are five main types of evaluation functions
(Dash & Liu, 1997):
1. Distance (Euclidean distance measure)
2. Information (Entropy, information gain etc.)
3. Dependency (Correlation coefficient)
4. Consistency (Minimum features bias)
5. Classifier Error Rate (based on the classification algorithm)
78
The limitation of all the methods listed above is that they may lead to the selection
of a redundant subset of features. Chen and Lin (2006) indicated that variables
that are independently and identically distributed are not truly redundant. Noise
reduction and better class separation can be obtained by adding variables that are
presumably redundant. Chen and Lin (2006) reportedly indicated that a variable
which is completely useless by itself, can provide a significant improvement in the
performance when it is considered with other variables. In other words, two
variables that are useless by themselves can be useful together. Thus, selecting
subsets of variables together can provide good prediction results, as opposed to
ranking the variables according to their individual predictive power. Thus, the
approach proposed in this research applies the technique developed by Chen and
Lin (2006), which uses F-score and the Random Forest (RF) with SVMs for the
purpose of feature selection. This approach is discussed in detail in Section 4.4 of
this thesis.
3.3.5 Classification
During the classification stage, the patterns of the ROIs (abnormalities in breast
tissue) are classified as either benign or malignant on the basis of the optimum
subset texture features selected for machine learning modeling. A classifier trained
on known abnormalities (mass lesions and microcalcifications/MCCs) combines
the selected features and uses confidence measures to indicate that a ROI is either
malignant or benign (Duda et al., 2001), (Bishop, 1995), (Ripley, 1996), (Fukunaga,
1990), (Vapnik, 1995), (Vapnik 1998).
During the training stage of learning machines, techniques to evaluate the learning
and memorization performance of a classification engine (classifier) are identified,
79
which include: Cross-validation (CV), leave-one-out scheme, as discussed in
Section 4.1.3.
Several automated classification techniques have been investigated for the
detection of mass lesions and MCCs in mammograms during the last decade. These
classification techniques include: Support Vector Machines (SVMs) (Wei et al.,
2005), Artificial Neural Networks (ANNs) (Wu et al., 1992), (Chan et al., 1995a),
(Jiang et al., 1996), (Sahiner et al., 1996), (Chan et al., 1997), (Huo et al., 1998),
(Papadopoulos et al., 2002), (Zhang et al., 2005), LDA (Chan et al., 1995b), (Zhang
et al., 2005), Convolutional Neural Networks (CNNs) (Sahiner et al., 1996) and the
k-Nearest neighbor (Veldkamp et al., 2000). Other techniques include a statistical
method based on the use of statistical models and the general framework of
Bayesian image analysis was developed by Karssemeijer (1993).
The authors in Yoshida et al. (1996) used the decimated wavelet transform and a
supervised machine learning technique for the detection of mass lesions and MCCs.
Suzuki et al. (2005) proposed a method for distinction between benign and
malignant tumors using ANNs. The authors in Cheng et al. (1998) proposed a fuzzy
logic based approach for classification between different breast abnormalities. The
authors in Hadjiiski et al. (1999) applied a hybrid combination of the Adaptive
Resonance Theory (ART) and LDA for classification between malignant and benign
abnormalities.
Recent studies have shown the superiority of SVMs over other supervised machine
learning techniques such as the ANN, suggesting that the SVM is a promising
technique for classification of noisy data. The authors in El-Naqa et al. (2002) used
80
SVMs to detect mass lesions and microcalcifications based on finite image
windows. Their approach relies on the capability of SVM to automatically learn
relevant features for optimal detection. In their work, a sensitivity of as high as 98
percent was achieved. Since then, SVMs have been proven to be useful for the
classification of masses and MCCs in digital mammography applications, as
indicated by the research groups, namely, Martins et al. (2009), Dehghan et al.
(2008), Papadopoulosa et al. (2005), Gorgel et al. (2009), Wei et al. (2005) and
Manzano-Lizcano et al., (2004). Thus, the approach proposed in this research
applies SVM for the purpose of pattern classification between malignant and
benign texture features. The theoretical background of SVM is discussed in detail
in Section 4.2 of this thesis.
3.3.6 Performance Evaluation
In order to evaluate the performance of pattern classification systems (developed
using supervised machine learning techniques such as ANNs and SVMs), the binary
classification performance has to be measured.
Table 3.1: Relation between, TP, TN, FP and FN ― Confusion matrix
Confusion Matrix Positive (#0) Negative ()0)
Positive (#1) True Positive (TP) False Positive (FP)
Negative ()1) False Negative (FN) True Negative (TN)
The performance of a binary classifier cannot be described by a single value and is
usually quantified by its accuracy during the test phase, i.e., the fraction of
misclassified points on the test set. The performance of a binary classifier can be
81
best described in terms of its sensitivity and specificity, quantifying its performance
to false positive (FP) and false negative (FN) instances (Veropoulos, 2001).
In a Receiver Operator Characteristics (ROC) curve, the sensitivity, which in this
research is the portion of malignant tumors that are correctly classified by the
learning machine, is plotted against 1-specificity, the share of benign tumors that
are falsely classified by the learning machine, for different cut-off values. The ROC
analysis generally is used to determine an optimum cut-off value referred to as
criterion for use in medical diagnostic tests. It is possible to achieve an optimal
balance between sensitivity and specificity that is needed for a certain purposes.
This can be achieved by changing the cut-off value of the system. Also, if the cost of
not detecting a particular disease becomes high to society, the cut-off value can be
changed to achieve a very high sensitivity, but lower specificity (Veropoulos,
2001).
Table 3.2: Binary classification performance measures
Performance Measure Definition
True Positive (TP) Tumor marked as malignant by a biopsy, which is also classified as malignant by the learning machine.
False Positive (FP) Tumor marked as malignant by a biopsy, which is classified as benign by the learning machine.
True Negative (TN) Tumor marked as benign by a biopsy, which is also classified as benign by the learning machine.
False Negative (FN) Tumor marked as benign by a biopsy, which is classified as malignant by the learning machine.
The confusion matrix in Table 3.1 indicates the relationship between different
performance indices for binary classification. For the computerized classification
82
of malignant and benign abnormalities in digital mammograms, the four
performance indices (TP, TN, FP and FN) in Table 3.2 are calculated by comparing
the predicted output from the learning machine with the real labels determined by
a biopsy. Using these four performance measures, relative measurements for
binary classification can be calculated. Sensitivity is defined as the ratio of tumors
which are marked and classified as tumor, to all marked tumors, given by:
Sensitivity = Positivescorrectlyclassified
Totalpositives= TP
TP2FN (3.1)
Specificity is defined as the ratio of tumors which are not marked and also not
classified as tumor, to all unmarked tumors, given by:
Specificity =Negativescorrectlyclassified
Totalnegatives= TN
TN2FP (3.2)
The overall accuracy is the ratio between the total number of correctly classified
instances and the test set size, given by:
Accuracy = Instancescorrectlyclassified
Totalinstances= TP2TN
TP2TN2FP2FN (3.3)
In order to visualize ROC curves of the binary classification performance, the
performance metrics True Positive Fraction (TPF) and False Positive Fraction (FPF)
can be computed using:
TPF =�Sensitivity) = 45678987:6;5<<:;7=>classified?57@=A568789:6 (3.4)
FPF=�1-Specificity) = K:L@789:6;5<<:;7=>classified?57@=K:L@789:6 (3.5)
83
In a medical diagnosis test, sensitivity gives the percentage of correctly classified
diseased individuals and specificity indicates the percentage of correctly classified
individuals without the disease. So, ROC curves are two-dimensional
representations the relative tradeoff between the sensitivity (TPF) and the 1-
specificity (FPF) of medical diagnostic test (Veropoulos, 2001), as shown in Figure
3.3.
Figure 3.3: An ROC curve ― FPF vs. TPF
The total Area Under the Curve (AUC) of a ROC curve represented by MN, which is a
quantitative measure of the binary classification performance as it reflects the
testing performance of the classifier at all possible cut-off levels. The larger the
AUC is within the closed interval [0.5,1], the better will be the classification
performance (Veropoulos, 2001).
There are generally a finite number of points on a ROC curve in most medical
diagnostic experiments, which mostly results in not finding a good approximation
of the AUC. Thus, the more points there are, the better will be the estimate of the
curve and the accuracy of the accuracy of the binary classifier.
84
(a) (b)
Figure 3.4: ROC curves (a) The likelihood of a tumor being benign or malignant
(b) The ROC curves of Figure 3.4 (Veropoulos, 2001)
There are several ways to calculate the AUC of a ROC curve. The trapezoidal rule
can be used to calculate the AUC, but it generally gives an underestimate of the
area. Another way to get a better approximation of the AUC is by fitting the TPF
and FPF data (in equations (3.4) and (3.5) respectively) into a binomial model
using curve-fitting software (Veropoulos, 2001). Sample distributions of benign
and malignant tumors visualized using ROC curves are shown in Figure 3.4.
The horizontal axis represents the certainty level that a tumor is malignant. When
the system has difficulty in identifying if a tumor is malignant or benign, the two
distributions overlap, as shown in Curve A (see Figure 3.4(a)). The AUC of Curve A
is 0.5 (see Figure 3.4(b)), is the worst performance that can be obtained. Curve C in
Figure 3.4(a) has the smallest overlap between the malignant and benign portions
85
which results in a near perfect performance with an MN nearly equal 1.0 (see Figure
3.4(b)).
Being such a useful performance graphing method, ROC curves have been rapidly
applied in several research areas, such as: medical decision-making (Veropoulos,
2001), machine learning and data mining (Spackman, 1989). In particular, ROC
curves have been in signal detection theory over the past few decades to depict the
tradeoff between benefits (TPs) and costs (FPs) of pattern classification systems
(Egan, 1975).
3.4 Fundamentals of Digital Image Processing
The knowledge of the concepts of digital image processing presented in the
following sections is required to fully comprehend the methods discussed in this
thesis. In particular, emphasis is given on the image processing and segmentation
techniques, which have been identified by the literature reviewed in Sections 3.3.1
and 3.3.2.
3.4.1 Representation of a Digital Image
Digital images are usually stored in a matrix form and represented as two-
dimensional functions of space. Let ��O, P� be this function representing an image.
Intensity function and Impulse function are some of the names used to denote ��. �. Parameters O and P are respective row and column coordinates of a pixel in the
image matrix and the value of ��. � is the intensity value of a pixel. The terms gray
levels and intensity levels both imply the value of a pixel in a grayscale image and
are used interchangeably in this thesis.
86
Figure 3.5: Coordinate convention used to represent digital images
(Gonzalez and Woods, 2002)
The coordinate convention that is widely used to represent digital images is shown
in Figure 3.5. According to the notation explained above, a digital image can be
represented mathematically in a matrix form as follows:
��O,P� =QRRS ��0,1� ��0,2� … ��0,U���1,1� ��1,2� … ��1,U�⋮ ⋮ ⋮��W, 1� ��W, 2� … ��W,U�XY
YZ , (3.6)
where it is assumed that the image has W rows and U columns, or in other words,
the image consists of W ×U pixels, where each element of the matrix ��O, P� corresponds to an image pixel. As indicated from Figure 3.5, that unlike Cartesian
coordinates, O is the vertical axis and P is the horizontal axis of a digital image. All
mammographic images acquired in this thesis for computer-aided modeling and
experimental work consists of 1024 × 1024 pixels as indicated in Section 5.2.1.
87
(a) (b)
Figure 3.6: Grayscale image (a) A dummy matrix A (b) Image corresponding
to matrix A
3.4.1.1 Range of Intensity Values
The range of intensity values for a digital image implies a closed interval
��\�, ��0]�, where ��\� is the smallest intensity value an image pixel can attain and
��0] is the largest intensity value. The size of the intensity range of images during
an image processing task is decided in accordance with the specifications and
requirements of the task. It is a common practice to assign brighter shades to
higher intensity values so that the pixel with highest value has a white shade and
the pixel with lowest value has black (darker) shade. An example of an image
matrix M and its corresponding gray scale image is illustrated in Figure 3.6.
When it comes to intensity range, it might seem that there are no limitations other
than that intensity values can only be real values. This might be true in theory;
however, due to the process involved in the acquisition of digital images, hardware
considerations and other factors, the number of intensity values is typically an
integer power of 2. Assuming that pixel values of an image can have ^ discrete gray
values in the range 0, ^ − 1�, and the allowed pixel values are equally spaced in
this interval, then there is a positive � such that (Gonzalez & Woods, 2002):
88
^ = 2_ (3.7)
The digital mammography images acquired in this research (for the development
of a framework for the computerized detection of breast cancer), have ^ = 256
gray values in the closed interval 0,255�, as discussed in Section 5.2.1. If a
mammogram image is of dimension W ×U then the number of bits required for
storage can be represented by W ×U × �. A special case occurs when � = 1, which
implies only two discrete intensity values allowed for a pixel. Such an image is also
referred to as a binary image or a black and white image. A binary image is usually
a result of the image segmentation process. In a binary image, as common practice
indicates, the white shade is assigned to pixels that are objects of interest and the
black shade is assigned to the rest of the pixels (background). For further details
about image segmentation, refer to Section 3.4.3 of this chapter.
3.4.2 Histogram
A histogram is a discrete function that describes occurrence of different gray levels
in an image. If the intensity range of an image `�O, P� is 0, ^ − 1�, then a histogram
can be defined as a functiona�b_� = )_ , where )_ is the number of pixels that have
�cd gray level b_ (Gonzalez & Woods, 2002). It is then possible to modify matrix A
in Figure 3.6(a) in to matrix B such that:
Figure 3.7: A dummy matrix �
89
The grayscale image in Figure 3.8(a) corresponds to matrix � in Figure 3.7 where
Figure 3.8(b) represents the corresponding histogram of matrix �. As observed
from Figure 3.8(b), gray level 0 occurs 2 times, gray level 1 occurs 12 times, gray
level 2 occurs 6 times, and so on.
3.4.2.1 Uses of Histogram
A histogram can be used to calculate and derive different properties from an
image. Most of the image processing methods based on histogram analysis are
statistical in nature and are related to the probability distribution of intensity
values. In this thesis it has been observed that a major advantage of performing
calculations on a histogram, instead of the image, is time complexity. Operations on
large images will take a considerable amount of time especially if performed
without optimization. On the other hand, a histogram will provide two 1D arrays,
each of length ^ for a �-bit image as given in equation (3.7). One of these arrays
contains the discrete gray levels and the other array contains the corresponding
values denoting the occurrence of these gray values. Even if the images are larger
than the above mentioned size, arrays obtained from the histogram are still of the
same size of length ^. Moreover, due to the single dimension of both these arrays,
operations are relatively simple.
Even though histogram calculations may seem relatively easy and simple, there are
certain drawbacks to be considered. Before being able to utilize a histogram, one
has to calculate it. Calculation of an image histogram can be a time consuming
operation for very large images. It is also shown in this research that methods
based alone on histograms are not adequate for the primary goal of this research.
Last but not the least, a histogram provides an observer with the so called global
90
information about an image. It is not possible to extract any local features from the
histogram of an image.
(a) (b)
Figure 3.8: Grayscale image. (a) Grayscale image corresponding to matrix B in
Figure 3.7 (b) Image histogram of grayscale image in Figure 3.8(a)
3.4.2.2 Histogram Normalization
Sometimes it is useful to normalize an image histogram by dividing each of its
values by the total number of pixels in the image. This calculation is also referred
to as histogram normalization, which creates a scaled histogram represented by
#�b_� = )_/), where ) is the total number of pixels in the image. From a statistical
point of view, # contains the probability distributions of different intensity values
of the image.
3.4.2.3 Histogram Equalization
Histogram equalization is an interesting image enhancement tool. Figure 3.9 shows
an example of histogram equalization on a grayscale image. As it is observed,
Figure 3.9(a) represents an image of pollen with poor contrast and Figure 3.9(b)
91
corresponds to its histogram that shows a high concentration of dark intensity
pixels. If a visual analysis of the image in Figure 3.9(a) is required, it would be
desirable to enhance the contrast in this image. Contrast enhancement can be
achieved by applying a transformation function so that the intensity distribution
becomes uniform in nature as indicated by the image histogram in Figure 3.9(d).
Such transformation is called histogram equalization, where the corresponding
grayscale image (in Figure 3.9(c)) has a brighter contrast with a balanced
distribution of white and dark intensity pixels.
3.4.3 Image Segmentation
One of the most recurrent prerequisites of an image processing system is the
ability to analyze images and detect regions that have specific characteristics. For
instance, there is a huge demand of techniques that can enable computers to
extract abnormalities and other malformations from human tissue. Such regions, in
general, are called Regions of Interest (ROIs) for obvious reasons.
Image segmentation is a process that divides image pixels into smaller structural
units that correspond to a ROI or neighborhood. Several segmentation algorithms
are available today and the performance of these segmentation algorithms is goal
specific. As explained by Rangayyan (2005), image segmentation techniques can be
classified into three major categories:
1. Thresholding techniques.
2. Boundary-based methods.
3. Region-based methods.
92
(a) (b)
(c) (d)
Figure 3.9: Illustration effects of histogram equalization
(Gonzalez and Woods, 2002)
The following sections discuss in detail each segmentation technique and Section
3.4.3.3.1 identifies the most suitable technique.
3.4.3.1 Thresholding Techniques
The main principle behind thresholding is that image pixels, falling into a
predefined range of intensity values, are assigned a single intensity value, and the
93
remaining pixels are assigned a different intensity value. A thresholding function
can be formally defined as,
*�O, P� = f& &� ��O, P� ∈ ^h , ^i�j &� ��O, P� ∈ ^h , ^i� , (3.8)
where *�. � is the thresholded version of image ��. � in equation (3.6), & and j are
the two intensities used to differentiate between two groups of pixels, and ^h and
^i are the lower and upper limits of the intensity range used to define the two
groups (Gonzalez & Woods, 2002). Definition of the intensity range varies
depending on the images and task in hand. One can simply define a single intensity
value as a threshold or define several intensity ranges. The resultant image only
contains two pixel intensities, representing a binary image. For simplification
purposes, in a binary image it is common to use intensity values 0 (black) and 1
(white) for & and j respectively.
(a) (b)
Figure 3.10: Segmentation using thresholding techniques. (a) Image from the
MIAS database (Suckling et al., 1994). Notice that the pectoral muscle has been
removed to show the effects of thresholding on glandular tissue only (b)
Thresholded image (Figure 3.10(a)) with a threshold value of 165
94
The basic principle in thresholding makes it highly suitable for segmenting ROIs
that will have distinct intensities from the rest of the image. An example image
taken from the Mammography Image Analysis Society (MIAS) database (Suckling
et al., 1994) is shown in Figure 3.10(a). The mammogram image in Figure 3.10(a)
is segmented by choosing the intensity value 165 as a threshold, which as a result
produces Figure 3.10(b) in which the white pixels correspond to the glandular
(breast) tissue. Thus, the ROI of the image is the brighter glandular breast shown
in Figure 3.10(b). The type of thresholding illustrated in Figure 3.10 is referred to
as global thresholding, where a single threshold intensity is used to segment the
grayscale image. However encouraging this result might seem, there are vital parts
of the glandular tissue that have not been included in the segmentation. This is due
to the fact that the type of thresholding explained here has limitations and requires
additional image processing to enhance the quality of the segmentation
(Rangayyan, 2005).
3.4.3.2 Boundary-based Methods
In terms of image processing, boundary in an image can be defined as a single
continuous edge forming a closed path and thereby enclosing a part of the image
that might be considered a ROI. A digital edge is formed by a region where
transition between dark and light pixels is sharp. This phenomenon is illustrated in
Figure 3.11(a) which presents a special case where the transition is rather abrupt.
Boundaries made up of such an edge are easy to detect automatically, and the term
ideal edge is used to denote them.
In boundary detection, initially an image is scanned for very sharp intensity
transitions to detect edges with specific characteristics such as orientation, degree
95
of blurring, length, etc. Then edge-linking algorithms are applied to create enclosed
boundaries. Edge-linking and boundary detection methods are explained in detail
in Gonzalez and Woods (2002). Furthermore, Rangayyan (2005) and Gonzalez and
Woods (2002) point out that alternative edge detection methods such as the
Hough-transform based global transformation and global processing via graph
theoretic techniques provide better results when it comes to creating an enclosed
boundary from several disjoint sets of edge pixels.
(a) (b)
(c) (d)
Figure 3.11: Example of an ideal edge and a blurred edge
(Gonzalez and Woods, 2002)
It is rarely the case when all ROIs are enclosed by an ideal boundary; this makes
boundary detection more complex. In reality an edge will look more like in Figure
3.11(b) which is a mammogram image acquired from the MIAS database (Suckling
96
et al., 1994). Optics, sampling and imperfections during image acquisition are some
of the reasons why edges are blurred (Gonzalez & Woods, 2002). The intensity
plots of the images in Figures 3.11(c) and 3.10(d) show the transition curve of the
intensity values (Gonzalez & Woods, 2002).
3.4.3.3 Region-based Methods
In most cases, regions that are to be extracted from an image have a specific
texture that is different from other parts of the image. Thresholding and boundary
detection might not be able to recognize specific texture because it is a local
property; this is when region-based methods can be applied. In region based
methods, it is assumed that pixels in a certain ROI with a specific texture share
similar characteristics. A cluster of pixels sharing similar values is referred to as
neighborhood in region-based methods. According to Rangayyan (2005) there are
two types of region-based segmentation: region splitting and merging and region
growing.
In region splitting and merging, an image is subdivided into smaller regions until
some predefined conditions are fulfilled. For instance, a condition might be that a
region should not be split if all of its pixels have the same intensity or have same
fractal dimension. Then the smaller regions are merged according to some pre-
specified conditions. This process might be continued until the desired result is
achieved.
Region growing methods usually start with a very small group of pixels and grow a
region by connecting neighborhood pixels that possess similar properties.
Different properties can lead to different regions. The initial starting point of
97
region growing is called a seed pixel. It is important to choose a correct seed pixel
to get the desired result. The choice of selecting a seed pixel can depend on several
task specific conditions and prior assumptions. For instance, if an approximate
location of the ROI is known, then the spatial centroid of this region can be used as
the seed pixel. Region-based methods are often time consuming, and in some cases
it is difficult to define the optimal conditions that will lead to the desired
segmentation.
3.4.3.3.1 Selected Segmentation Technique
Segmentation of mammograms for the computerized detection of breast cancer
can be typically performed using the three approaches discussed in Section 3.4.3.
From the literature reviewed in Section 3.4.3, this section identifies the most
suitable segmentation technique for the purpose of mammogram segmentation.
Thresholding techniques are based on the criterion that all pixels in an image
whose pixel intensities lie within a specific range belong to a particular class (or
neighborhood). Typically, thresholding methods omit all the spatial information of
an image and do not deal well with noise. Similarly, boundary-based methods use
the criterion that intensity values of pixels in an image change quickly at the
boundary between the two regions. The basic approach used in boundary-based
methods is to apply a gradient operator such as the Sobel filter (Ballard & Brown,
1982). High values of such a filter typically give candidates of region boundaries.
The candidates of these boundaries need be modified so as to establish curves
corresponding to the boundaries between the connected regions. Conversion of
the edge pixels into boundaries of the ROIs is a challenging task. Working with
regions is the complement of boundary-based approaches.
98
Region-based approaches rely on the criterion that all the neighboring pixels
within one region have identical values (Horowitz and Pavlidis, 1974). In region-
based approaches, the general task is to compare the pixel of interest with its
connected neighborhood. If a criterion of homogeneity is satisfied in a region-
based approach, the pixel is said to belong to the same class (or region) as one or
more of its neighbors.
Segmentation techniques such as thresholding and boundary-based techniques are
not able to recognize specific textures because they consider texture regions as a
local property. Since the pectoral muscle in mammograms is represented as a
texture property of bright contrast intensity pixels, and needs to be suppressed
from the mammogram (Kwok et al., 2004), (Raba et al., 2005), (Nicolaou et al.,
2008), (Xu et al., 2007), (Ferrari et al., 2004), (Mirzaalian et al., 2007), (Bajger et
al., 2005), thus, region-based methods are applied in this research to suppress un-
important regions in mammogram images (Pavlidis & Liow, 1990). The theoretical
concepts of an advanced region-based technique, namely, Seeded Region Growing
(SRG), applied for pectoral muscle segmentation, is presented in the following
section.
3.4.3.3.2 Seeded Region Growing
Region growing (Cheevasuvit et al., 1986), (Pavlidis & Liow, 1990) is the most
commonly used region-based method for image segmentation. Seeded Region
Growing (SRG) is based on the traditional region growing criterion of the similarity
of pixels within regions (Adams and Bischof, 1994). Instead of optimizing
homogeneity parameters as the case with conventional region growing techniques,
SRG is operated by choosing a number of pixels, known as seeds, instead of tuning
99
homogeneity parameters as in traditional region growing. This type of control
allows unskilled and non-expert users to achieve good segmentation results on
their initial attempt.
In SRG, segmentation of an image is performed with respect to a set of points,
known as seeds. Consider a number of seeds grouped into ) sets, say, M�, M�,… , M� . In
some cases individual sets will consist of single points. The choice of seeds in SRG
decides that what are the features of interest and what are irrelevant or noise
(Chen et al., 1991). Given the seeds, the SRG tries to find a tessellation of the image
into smaller regions satisfying the condition that, each connected component of a
region meets exactly one of the M\; limited to this constraint, the regions are
selected to be as homogenous as possible. The description of the SRG method as
applied to grayscale images is presented below (Adams & Bischof, 1994). The SRG
process develops inductively from the choice of seeds selected, namely, the initial
state of the sets, M�, M�,… , M� . In SRG, each step of the algorithm performs addition of
one pixel to any of the above sets. Then considering the state of the sets M\ after k
steps, consider � be the set of all unallocated pixels, bordering at least one of the
regions such that (Adams & Bischof, 1994):
� = lO ∉ ⋃ M\|U�O� ∩ ⋃ M\ ≠ ∅�\s��\s� t (3.9)
where U�O� contains the immediate neighbours of the pixel of interest O. As an
example, the immediate neighbours are those pixels which are 8-connected to the
pixel of interest O. If for, O ∈ � we have that U�O� meets just one of the M\, then we
can define: (i) &�O� ∈ l1,2, … , )t to be that index such that U�O� ∩M\�]� ≠ ∅, and (ii) u�O�
100
to be a measure of how different O is from the region it joins. The simplest
definition for u�O� is (Adams & Bischof, 1994):
u�O� = v*�O� − meanz∈{|�}� *�P��v (3.10)
where *�O� is the grey level intensity of the image pixel O. If U�O� meets two or more
of the M\, &�O� is taken to be a value of & such that U�O� meets M\ and u�O� is also
minimized. In this circumstance, it is desirable to classify the pixel O as a boundary
pixel and append it to set �, which is a set of already-found boundary pixels. We
then take ~ ∈ � such that (Adams & Bischof, 1994):
u�O� = min]∈�lu�O�t (3.11)
and append ~ to M\�~�. This process completes step k+ 1. This entire process is
iteratively repeated until all pixels are allocated. In SRG, the process starts with
each M\ being one of the seed sets. Thus, the definitions of u�O� in equations (3.10)
and (3.11) ensure that the segmentation (result) is into regions as homogenous as
possible.
For implementing SRG using programming, a data structure termed as Sequentially
Sorted List (SSL) is used. A SSL contains a linked list of objects, which contains
pixel addresses that are ordered according to some feature or attribute. At the
beginning of each of step of SRG, when the algorithm considers a new pixel, the
pixel at the beginning of the list is taken out. When adding a pixel to the list, it is
placed according to the value of the ordering attribute. In the case of SRG, the SSL
stores the data of � (in equation (3.9)), which is ordered according to u�O�. The
101
algorithm for implementing SRG is presented in pseudo-code as follows (Adams &
Bischof, 1994):
Label seed points according to their initial grouping.
Put neighbors of seed points (the initial �) in the SSL.
While the SSL is not empty:
Remove the first point P from the SSL.
Test the neighbors of this point:
If all neighbors of P which are already labeled (other than with the
boundary label) have the same label, then:
SetP to this label.
Update running mean of the corresponding region.
Add neighbors of P which are neither already set nor already in the
SSL to the SSL according to their value of u.
Otherwise:
Flag P with the boundary label.
Based on the stepwise description shown above in the pseudo-code, it is observed
that in executing the SRG algorithm each pixel is visited once only, although at each
visit, each of the 8-connected neighboring pixels are also visited. Hence, this makes
SRG a rapid algorithm.
Achieving a good segmentation performance is dependent on choosing a correct
set of seeds as the homogeneity parameter (Cheevasuvit et al., 1986), (Chen et al.,
1991) and (Pavlidis & Liow, 1990). If the regions in grayscale image are noiseless,
the only thing necessary for a good segmentation performance using SRG is that
each pixel in a region should have a gray value which is similar to the mean of the
region. In most cases, if the regions have noise present, single seeds may fall on an
102
outlier, which can result in a poor starting estimate of the region’s mean, causing
the segmentation to be incorrect. In order to prevent this from happening, it is
recommended that large seed areas should be used when segmenting noisy
regions in images. Furthermore, the area of each seed should be large enough so as
to ensure that a stable estimate of its region’s mean can be obtained (Adams &
Bischof, 1994).
The choice of the region mean for the definition of u�O� is assumed such that, the
noise in each region contains equal variance. However, if this assumption is not
true, then an appropriate choice of u can be calculated using:
u�O� = ���]�/�:@��∈�|�}� ��z�����∈�|�}� ��z�� � (3.12)
where SD represents the sample standard deviation of the region. In SRG, if the SD
is a known function of the region’s mean, then a suitable technique for variance-
stabilizing should be applied to the image first.
3.4.4 Morphological Operations
Morphological operations used in digital image processing are a way of extracting
image components that can be used to express details about a regions shape, its
boundaries, its area and so on (Gonzalez & Woods, 2002). Two primitive and
widely used morphological operations dilation and erosion are discussed in the
following sections to fully comprehend the methods discussed in this thesis.
103
3.4.4.1 Dilation
Dilation is generally used to smooth boundaries of regions or bridge very small
gaps between neighboring regions. According to Gonzalez and Woods (2002), the
formal definition of dilation of a set A by another set B is denoted M⊕�, and
defined by:
M⊕� = �~|����N ∩ M ≠ 0� (3.13)
where �� is the reflection of �. This definition means that dilation of M by � is done
by reflecting � and then shifting � over M by ~. Then all the displacements of � are
set such that � and M overlap by at least one element, which gives the dilation. Set
� is also referred to as the dilation mask or structuring element (STREL). In Figure
3.12, an example of set M and a set � are shown to illustrate the effect of dilation.
The center of the mask � is marked by a small black square. In this case, the
reflection �� is equivalent to �. Now, if � is moved within and outside M, then
dilation is given by the set of all points traversed by the center of �, until M and �
are overlapped by at least one element. The resultant is shown as the shaded
square that is bigger in size than M as indicated by dashed lines.
Figure 3.12: An example of dilation of set A by set B
104
3.4.4.2 Erosion
Erosion produces an opposite effect of dilation. Following the same notation for
dilation in equation (3.13), a formal definition of erosion is given by (Gonzalez &
Woods, 2002):
M⊝� = l~|���N ⊆ Mt (3.14)
In other words, erosion of M by � is set of all points traversed by center of � such
that � is totally contained within M at all times. Figure 3.13 shows an example of
set M and a set � to illustrate the effect of erosion. Erosion can be used for
removing small unwanted components, such as thread like structures, from an
image by using a structuring element (STREL) with an area that is bigger than the
unwanted regions.
3.4.4.3 Morphological Opening and Closing
In digital image processing, the processes of morphological dilation and erosion
can be combined together in different ways to make interesting changes in images.
Morphological opening and closing are two such operations that are defined by
specific combinations of dilation and erosion. Morphological opening is generally
used to smooth region contours and remove thin protrusions in images. Similarly,
morphological closing adds smoothness to image contours; however, it generally
fuses two large regions separated by narrow breaks. This effect is opposite to that
caused by morphological opening that breaks the narrow links between two large
regions (Gonzalez & Woods, 2002).
105
Figure 3.13: An example of erosion of set A by set B
3.5 Texture Extraction and Analysis
The knowledge of the concepts texture extraction and analysis presented in the
following sections is required to fully comprehend the methods discussed in this
thesis. In particular, emphasis is given on the texture-based feature extraction
techniques and algorithms which have been identified by the literature reviewed
in Section 3.3.3.
3.5.1 Introduction to Texture Analysis
Texture analysis is an important area of research in computer vision and image
processing algorithms. The texture analysis domain composes of three major types
of problems, which include: texture classification, texture segmentation and
texture synthesis (typically used in image compression). Many definitions about
texture have been formulated by different computer vision researchers. The most
classical definition is as follows:
"Texture is defined for our purposes as an attribute of a field having
no components that appear enumerable. The phase relations between
the components are thus not apparent. Nor should the field contain
an obvious gradient. The intent of this definition is to direct attention
106
of the observer to the global properties of the display i.e., its overall
"coarseness," "bumpiness," or "fineness." Physically, non-numerable
(aperiodic) patterns are generated by stochastic as opposed to
deterministic processes. Perceptually, however, the set of all patterns
without obvious enumerable components will include many
deterministic (and even periodic) textures." (Richards and Polit,
1974).
(a) (b)
Figure 3.14: Textures (a) An image consisting of sixteen different textured regions
(b) Texture segmentation of (a) produced by an automatic procedure
There are three main ways of using texture features:
1. To segment an image based on texture (texture segmentation).
2. To discriminate between different (already segmented) regions or to
classify them into separate categories (texture classification).
3. To produce descriptions so that textures can be reproduced (texture
synthesis).
107
Conceptually, the main characteristic of a texture is basic pattern recurrence. The
structure of a pattern can be statistical and the pattern recurrence statistically
regular. Figure 3.14(a) shows an example of 16 texture categories identified as
separate textures using generic labels to classify texture patterns, based on the
different fill of the patterns as shown in Figure 3.14(b).
The general framework for the computerized detection of breast cancer in Figure
3.1 (Hutt, 1996) composes of a feature extraction and classification stage.
Computation of texture-based features for pattern classification provides cues for
classifying between patterns of malignant and benign abnormalities. As a basis for
all texture related applications, texture analysis seeks to produce a general,
efficient and compact quantitative description of textures so that mathematical
operations can be applied to alter, compare and transform textures. The following
sections discuss the most recent and successful texture-based approaches applied
to digital mammography.
3.5.2 Texture Analysis Applied to Digital Mammography
This section reviews some recent publications focusing on texture analysis used in
digital mammography applications and describes the contributions made. Recent
studies of breast cancer aim to improve the radiologist's diagnostic performance
by indicating suspicious areas. Juhl (1982) introduced the first study based on
specific features sought in routine examinations as common indicators of
malignancy, which supposes a big development and increment of investigation.
During the last decades Chan et al. (1990), Davies & Dance (1990), Karssemeijer
(1992) and many other research groups focused their work on microcalcification
detection. Other researches such as Giger et al. (1990) addressed their work in
108
investigating methods to detect and analyze mass lesions by means of asymmetry
studies. Moreover, it is important to remark that the study of Miller and Astley
(1992) introduced the concept of texture analysis used to classify between breast
tissues.
The increase of related works introduced by new techniques and improvements in
digital mammography are mainly due to the analysis of textures, which is
nowadays gaining much importance in breast cancer detection. Analysis of
textures in digital mammograms is beneficial since textures have been found
useful in identifying specific patterns of breast abnormalities. As digital
mammography produces high resolution gray level images where textures play an
important role, texture descriptors are required in order to select a set of
distinguishing and sufficient texture features for the purpose of characterizing
different textures in digital mammograms. Texture features contain information
about the spatial distribution on the intensity pixels within defined regions in a
grey scale image. Thus, the texture of a region of a digital mammogram describes
the pattern of the spatial variation of grey tones in a neighborhood. Gulsrud and
Loland (1996) presented an automated technique for the detection of different
abnormalities in digital mammograms based on the application of multichannel-
filtering for texture feature extraction.
There are several textural analysis techniques proposed in the literature for the
computerized detection of breast cancer detection in digital mammography
applications, however this thesis only presents the most recent and successful
approaches. The three most recent texture-based approaches presented in the
following sections are:
109
1. Grey-level Co-occurrence Matrices (GLCMs)
2. Law’s Textures
3. Local Binary Patterns (LBPs)
3.5.2.1 Gray-level Co-occurrence Matrices
Gray-Level Co-occurrence Matrices (GLCMs) are one of the recent and successful
texture analysis techniques commonly used for feature calculation in computer-
aided applications of digital mammography.
Bovis and Singh (2000) studied detecting masses in mammograms on the basis of
textural features using five GLCMs statistics extracted from four spatial
orientations, horizontal, left diagonal, vertical and right diagonal corresponding to
�0°, 45°, 90°and135°� and four pixel distance �� = 1,3,6and9�. Classification is
performed using each texture feature vector and the Linear Discriminant Analysis
(LDA). According to Martì et al. (2000), GLCMs have been frequently used in
computer vision and are known to obtain satisfactory results as texture classifiers
in different applications. Their approach uses mutual information with the purpose
to calculate the amount of mutual information between images using histograms
distributions obtained by GLCMs.
Blot and Zwiggelaar (2000a) proposed two approaches based on GLCMs for the
detection and enhancement of structures in images. In the first approach, the
GLCM of the local ROI is compared to a mean GLCM obtained from a number of
equal size areas surrounding the local ROI. The purpose is to compare the
difference between these two matrices obtaining a probability estimate of the
abnormal image structures in the ROI. In The second approach by Blot and
110
Zwiggelaar (2000b) follows the same proposal extracting background texture in
mammographic images with an improvement respect to previous works and it
does not depend on any prior assumptions about the type of structures to be
enhanced. Another study based on background texture extraction for classification
was presented by Blot and Zwiggelaar (2001), which indicated that there is a
statistical difference between the GLCM for image regions that include the image
structures and regions that only contain background texture. Thus, the
classification of mammographic parenchymal patterns can be improved if
anatomical structures can be removed from the image.
In 2003, different approaches of GLCMs for texture-based analysis of digital
mammograms were proposed. Youssry et al. (2003) presented a neuro-fuzzy
model for fast detection of candidate circumscribed masses in mammograms,
where texture features estimated using GLCMs are used to train the neuro-fuzzy
model. Similarly, Martì et al. (2003) proposed a supervised method for the
segmentation of masses in mammographic images using GLCM texture features
which present a homogeneous behavior inside selected regions. Jirari (2005)
proposed a computer-aided detection system for breast cancer detection using five
GLCMs at different distances for each suspicious ROI and used the texture
descriptors to train and test a Radial Basis Function Neural Network (RBFNN).
The authors in Costaridou et al. (2005) investigated how to differentiate dense
breast regions containing spiculated masses from regions of normal dense tissues,
by means of feature analysis on wavelet-processed mammograms. In this work
Costaridou et al. (2005) extracted multi-resolution texture features of second
order statistics from spatial GLCMs using different orientations and distances.
111
In more recent studies, Karahaliou et al. (2007) investigated texture properties of
the tissues surrounding microcalcifications using a wavelet spatially adaptive
enhancement method, namely the Wavelet transform. For this purpose the authors
in Karahaliou et al. (2007) calculated thirteen textural features from four GLCMs.
More recently, Lyra et al. (2008) investigated how to identify breast tissue quality
data quantification using a computer-aided detection, where images were
categorized using the BI-RADS lexicon indicated in Section 2.6.4.1. In this work,
Lyra et al. (2008) derived texture features for each sub-region using an averaged
GLCM. Karahaliou et al. (2008) investigated texture properties of the tissues
surrounding MCCs and mass lesions in digital mammograms using gray-level
texture and wavelet coefficient texture features at three decomposition levels. In
modern times, Khuzi et al. (2009) presented a technique to identify mass lesions in
digital mammograms using GLCM based texture descriptors.
3.5.2.2 Law’s Texture Filter
Previous studies of Laws' textures by Miller and Astley (1992) proposed an
approach to discriminate between glandular and fatty regions of breast tissue in
order to automate the detection of breast asymmetries. Karssemeijer (1992, 1993)
used Law’s masks as a mechanism for detecting architectural distortions caused
predominantly due to the ductal patterns of mammograms. Gupta and Undrill
(1995) applied Law’s masks to the task of delineating suspicious mass lesions and
examined whether this texture-based approach indicates prospects of
discrimination between stellate lesions and regular masses.
Another approach proposed by Pfisterer and Aghdasi (1998) investigated how to
detect masses in digitized mammograms using textural information. In this work,
112
enhancement and segmentation of the images is based on texture analysis, which is
performed using wavelets, steerable filters and Law’s texture maps (Pfisterer &
Aghdasi, 1998).
Bovis and Singh (2000) investigated a new approach for classification of
mammographic images according to the breast type using Law’s texture masks. In
this work, Bovis and Singh (2000) extracted the total texture energy for this mask
combination, for use as a feature. More work was presented by Varela et al. (2001)
based on a digital image processing algorithm to classify mass lesions based on
quantitative measures of tumor shape, contrast, and spiculation. Features based on
Law’s texture energy were extracted from the straightened border regions. Three
Laws’ filters (vertical, horizontal and symmetrical) were applied to the
transformed image (Varela et al., 2001). In the same year, Pfisterer and Aghdasi
(2001) following their previous work (Pfisterer & Aghdasi, 1998) continued their
study and found better results than a specific convolutional mask, because it is not
as important as the pre-enhancement used.
Karahaliou et al. (2006) demonstrated how texture analysis of a breast tissue
surrounding microcalcifications and mass lesions shows promising results
contributing to the reduction of benign biopsies. In this work, Karahaliou et al.
(2006) computed texture features from the remaining ROI area (surrounding
tissue) by using first and second order statistics algorithms, grey level run length
matrices and Law’s texture energy measures. Another more recent study of
Karahaliou et al. (2007) investigated the texture properties of the tissues
surrounding microcalcifications and masses using a wavelet spatially adaptive
enhancement method (wavelet transform). In this work, Karahaliou et al. (2007)
113
used Law’s texture energy measures for the discrimination of malignant from
benign tissues were investigated using a k-Nearest Neighbor classifier.
3.5.2.3 Local Binary Patterns
Local Binary Patterns (LBPs) is a recent and promising technique for texture
analysis. Oliver et al. (2007a) presented an approach to represent salient micro-
patterns using the spatial structure of masses. The approach proposed by the
authors in Oliver et al. (2007a) focuses on reducing the number of FPs in the field
of lesion detection and distinguishes between the true recognized masses. Another
work presented by Oliver et al. (2007b) is based on classifying mammograms
using texture information. In this sense, texture descriptors are extracted from
each cluster by using LBPs.
More recently, Lladó et al. (2007) analyzed a set of FP reduction methods in the
field of mammographic mass detection using different approaches to extract lesion
image features using LBPs. In modern times, Lladó et al. (2009) proposed the use
of LBPs to characterize micro-patterns and preserve the spatial structure of
masses.
3.5.3 Comparison of Texture Analysis Techniques
The literature reviewed on texture analysis techniques applied for digital
mammography in Section 3.5.2 indicates that GLCMs, Law’s texture filter and LPBs
are the most recent texture-based approaches applied to digital mammography.
Table 3.3 summarizes these textural analysis techniques with respect to the
majority of work conducted and success achieved by each. As observed from Table
114
3.3, in modern times, GLCMs have received a notable success for computer-aided
detection in digital mammography compared to Law’s texture filter and LBPs.
Table 3.3: Classification of the state-of-the-art texture analysis techniques
Year Author(s) Feature Extraction Technique
GLCM LAWS LBP
1992 Miller and Astley (2000)
Karssemeijer (1992) ���� ����
���� ����
���� ����
1993 Karssemeijer (1993) ���� ���� ����
1995 Gupta and Undrill (1995) ���� ���� ����
1998 Pisterer and Aghdasi (1998) ���� ���� ����
2000
Bovis and Singh (2000) Martì et al. (2000)
Blot and Zwiggelaar (2000a) Blot et al. (2000b)
Mudigonda et al. (2000)
���� ���� ���� ���� ����
���� ���� ���� ���� ����
���� ���� ���� ���� ����
2001
Varela et al. (2001) Pfisterer et al. (2001)
Blot and Zwiggelaar (2001)
���� ���� ����
���� ���� ����
���� ���� ����
2002 Blot et al. (2003)
Bovis and Singh (2002) ���� ����
���� ����
���� ����
2003 Youssry et al. (2003)
Martì et al. (2003) ���� ����
���� ����
���� ����
2005 Jirari (2005)
Costaridou et al. (2005) Oliver et al. (2005)
���� ���� ����
���� ���� ����
���� ���� ����
2006 Hassanien and Slezak (2006)
Karahaliou et al. (2006) ���� ����
���� ����
���� ����
2007
Oliver et al. (2007a) Karahaliou et al. (2007)
Lladó et al. (2007) Oliver et al. (2007b)
���� ���� ���� ����
���� ���� ���� ����
���� ���� ���� ����
2008 Lyra et al. (2008)
Howard et al. (2008) ���� ����
���� ����
���� ����
2009 Lladó et al. (2009) Khuzi et al. (2009)
���� ����
���� ����
���� ����
Some authors such as Karahaliou et al. (2007) and Oliver et al. (2007a) have
applied their approaches using two texture-based techniques, as indicated in Table
3.3. LBP is a modern and promising technique but there are still few studies where
it is applied in digital mammography. Lladó et al. (2007, 2009) and Olivier et al.
115
(2007a, 2007b) are the most recent researchers to apply LBPs on digital
mammograms. As observed from Table 3.3, previous studies on digital
mammography emphasizes Law’s texture filter, however in modern times GLCMs
have been chosen to replace Law’s texture filter (Khuzi et al., 2009).
3.5.3.1 Selected Texture Extraction Technique
Nowadays many techniques in computer vision deal with feature extraction in
image processing. Texture features and statistical features are the most significant
heuristics for the purpose of pattern recognition from digital images. Frequently
used approaches for texture analysis are mainly based on statistical properties of
the intensity histogram. Commonly used texture based feature extraction
techniques include: autocorrelation function of textures GLCMs, Fractal texture
description, Law’s texture filter, LPBs etc. Many statistical feature extraction
methods are present in literature, the most common include: edge frequency,
primitive length (run length), mathematical morphology, Gabor transform,
wavelets, etc. Other feature extraction methods are geometrical methods known as
texton, which analyze the structure of textures by identifying basis elements. For a
complete description and classification of texture analysis, refer to Gonzalez and
Woods (2002).
GLCMs have received a notable success in recent years (see Table 3.3) for the
computerized detection of breast cancer in digital mammography applications
compared to Law’s texture filter and LBPs. With the increasing number of
publications on the application of GLCMs for digital mammography, it is easy to
think about the importance of GLCMs compared to the other texture analysis
116
techniques. Thus, this research applies GLCMs for texture-based feature extraction
from digital mammograms using texture descriptors discussed in (Haralick et al.,
1973), (Soh and Tsatsoulis, 1999) and (Clausi, 2002). The following section
presents the background and theoretical concepts of GLCMs.
3.5.3.1.1 Introduction to GLCMs
The statistics of grey-level histograms give parameters for each processed region,
but do not provide any information about the repeating nature of the texture.
According to Beichel and Sonka (2006), the occurrence of gray-level configuration
may be described by matrices of relative frequencies, called co-occurrence
matrices. Hence, the Grey-level Co-occurrence Matrix (GLCM) is a tabulation of how
often different combinations of pixel intensity values (grey levels) occur in an
image. GLCMs are constructed by observing pairs of image cells at a distance �
from each other and incrementing the matrix position corresponding to the grey
level of both cells. This allows deriving four matrices for each given the distance:
��0°, ��, ��45°, ��, ��90°, ��, ��135°, ��. For instance, ��0°, �� is defined as
follows:
� ��0°�, ��+, ��� = �����, ���k. )�� ∈ �: � −k = 0, |� − )| = �,���, �� = +, ��k, )� = � �� (3.15)
where each � value is the number of times that: ��O�, P�� = &, ��O�, P�� =j, |O� − O�| = � and P� = P� append simultaneously in the image. Similarly,
��45°, ��, ��90°, ��and ��135°, �� can be defined as follows:
117
� ��45°�, ��+, ��� = �����, ���k. )�� ∈ �: �� − k = �, |� − )| = −��OR�� − k = �, |� − )| = ��,���, �� = +, ��k, )� = � ¢� (3.16)
� ��90°�, ��+, ��� = �����, ���k. )�� ∈ �: |� − k| = �, � − ) = 0,���, �� = +, ��k, )� = � �� (3.17)
� ��135°�, ��+, ��� = �����, ���k. )�� ∈ �: �� − k = �, � − ) = ��OR�� −k = −�, � − ) = −��,���, �� = +, ��k, )� = � ¢� (3.18)
A GLCM contains the frequency of a certain pair of pixels repetition in an image.
According to equations (3.15) to (3.18), the parameters required for computing
GLCMs are as follows:
• Number of Grey Levels: Normally, GLCMs use a grayscale image of 256
grey levels, which indicates a higher computational cost because all
possible pixel pairs must be taken into account. The solution is to generate
the matrix reducing the number of grayscale, and so the number of possible
pixel combinations. The GLCM is always square with the same
dimensionality as the number of grey-levels chosen.
• Distance between Pixels (£): GLCMs store the number of times that a
certain pair of pixels is found in an image. Normally the pair of pixels are
neighbors, but the matrix could also be computed analyzing the relation
between non-consecutive pixels. Thus a distance between pixels must be
previously defined.
118
• Angle (¤): Similar to the distance parameter, it is necessary to define the
direction of the pair of pixels (neighbors). The most common directions are
0°, 45°, 90°, 135° and its symmetric equivalents.
Figure 3.15: Spatial relationships of pixels defined by offsets, where � is
the distance from the pixel of interest
Figure 3.15 illustrates the spatial relationships of pixels defined by the two
parameters: the distance between the pixels ��� and the angle �¥�. In order to
illustrate the process of computation of GLCMs, Figure 3.16 shows the process
used to create GLCMs. The matrix on the left in Figure 3.16 is the input image and
matrix � on the right is the GLCM of the input image.
Figure 3.16: Process used to create GLCMs
119
As an example, consider calculating the first three values of matrix � of Figure
3.16. In matrix �, element�1,1� contains the value 1 because there is only one
instance in the input image where two horizontally adjacent pixels have the values
1 and 1 respectively. Similarly, element�1,2� in matrix � contains the value 2 as
there are only two instances where two horizontally adjacent pixels in the GLCM
have the values 1 and 2. In matrix �, element�1,3� has the value 0 as there are no
instances of two horizontally adjacent pixels with the values 1 and 3 in the GLCM.
The GLCM scans the rest of the input image for other pixel pairs �&, j� and the sums
are recorded in the corresponding elements of the GLCM.
3.5.3.1.2 GLCM Texture Descriptors
GLCMs contain the information and properties about the spatial distribution of
gray levels in grayscale images. The basis for GLCM texture descriptors comes from
the GLCM. The GLCM is a square matrix with dimension U�, where U� is the
number of gray levels in the grayscale image, represented in equation (3.19).
Element #�&, j� of matrix G (in equation (3.19)) is generated by first counting the
number of times a pixel with value & is adjacent to a pixel with value j in the GLCM
and then by dividing the entire matrix by the total number of comparisons made.
Thus, each entry is considered to be the probability that a pixel with value & will be
adjacent to a pixel with value j.
G =QRRRS#�1,1� #�1,2� ⋯ #�1, U��#�2,1� #�2,2� ⋯ #�2, U��⋮ ⋮ ⋱ ⋮#�U� , 1� #�U� , 2� ⋯ #�U� , U��XY
YYZ (3.19)
120
Since adjacency can occur in each of the four directions (i.e., horizontal, vertical
and left and right diagonals) for a two-dimensional image as shown in Figure 3.15,
four matrices, each corresponding to one direction can computed as given in
equations (3.15) to (3.18).
Haralick et al. (1979) described fourteen texture descriptors for characterizing
GLCMs, which are shown in Table 3.4 (Haralick et al., 1979) in equations (3.20) to
(3.33). The texture descriptors proposed by Haralick (1973) form the basis for
texture feature extraction using GLCMs. The term #�&, j� in Table 3.4 represents the
&jcd term of matrix � (in Figure 3.16) divided by the sum of the elements of � . In
this research apart from using texture descriptors proposed by Haralick et al.
(1979), other recent texture descriptors proposed by Soh & Tsatsoulis (1999),
Clausi (2002) and the MATLAB Image Processing Toolbox are evaluated, as
indicated in Section 5.4.2 of this thesis.
3.6 Summary
This chapter presented and discussed the background literature on the computer
processing of digital mammograms. In Section 3.1, computer-aided detection
systems were introduced with the literature review of computerized breast cancer
detection techniques presented in Section 3.2. Section 3.3 identified the key
techniques and algorithms used in this research to develop a framework for the
computerized detection of breast cancer. Section 3.4 presented the fundamental
concepts of digital image processing with emphasis on image segmentation
techniques used in digital mammography applications. Lastly, Section 3.5
emphasized on the use of texture-based analysis for the purpose of feature
extraction in pattern classification problems.
121
Table 3.4: Standard GLCM texture descriptors (Haralick, 1973)
No. Texture Descriptor Formula Equation No.
1. Angular Second Moment: Energy
∑ ∑ l#�&, j�t�ª\ (3.20)
2. Contrast ∑ )�«¬/��s �∑ ∑ #�&, j�, |& − j| = )«¬ªs�«¬\s� � (3.21)
3. Correlation (Haralick)
∑ ∑ �\ª���\,ª�/®}®�¯}¯�ª\
where, °] , °z , ±] and ±z are the means and standard deviations of #] and #z , the partial probability density functions.
(3.22)
4. Sum of Squares: Variance
∑ ∑ �& − °��#�&, j�ª\ (3.23)
5. Inverse Difference Moment
∑ ∑ ��\,ª��2�\/ª�²ª\ (3.24)
6. Sum Average
∑ &#]2z�&��«¬\s�
where, O and P are the coordinates (row and column) of any entry in the co-occurrence matrix, and #]2z�&� is the probability of co-occurrence matrix coordinates summing to O + P.
(3.25)
7. Sum Variance ∑ �& − ³�c��#]2z�&��«¬\s� (3.26)
8. Sum Entropy −∑ #]2z�&�log·#]2z�&�¸�«¬\s� = ³�c (3.27)
9. Entropy −∑ ∑ #�&, j�log�#�&, j��ª\ (3.28)
10. Difference Variance ∑ &�#]/z�&�«¬/�\s (3.29)
11. Difference Entropy −∑ #]/z�&�«¬/�\s log·#]/z�&�¸ (3.30)
12. Information Measure of Correlation 1
¹º» − ¹º»1maxl¹º,¹»t where, ¹º and ¹» are the entropies of #] and #z such that: ¹º» = −∑ ∑ #�&, j�log�#�&, j��ª\ ¹º»1 = −∑ ∑ #�&, j�log·#]�&�#z�j�¸ª\ ¹º»2 = −∑ ∑ #]�&�#z�j�ª\ log·#]�&�#z�j�¸
(3.31)
13. Information Measure of Correlation 2
�1 − exp −2�¹º»2− ¹º»���� �⁄ (3.32)
14. Maximum Correlation Coefficient
¾�SecondlargesteigenvalueofÃ� where, Ã�&, j� = ∑ ��\,_���ª,_��}�\����_�_
(3.33)
122
CHAPTER 4
PATTERN RECOGNITION AND FEATURE SELECTION
4.0 Overview
An overview of pattern recognition is given in this chapter, with particular
emphasis on a specific machine learning technique, namely, the Support Vector
Machine (SVM). SVMs will be used intensively in this research. The reason for
using SVM as the main machine learning technique for this research is discussed in
Sections 1.3 and 3.3.5. Section 4.1 presents some introductory notions regarding
the theoretical concepts of learning machines. Section 4.2 introduces the
fundamental concepts of the statistical learning theory and presents the
mathematical formulation of the SVM developed by Vapnik (1998) which describes
the statistical aspects of automated machine learning. Towards the end of this
chapter, Section 4.3 presents the theoretical concepts of ANNs whereas Section 4.4
discusses a Recursive Feature Elimination (RFE) technique used for the selection
of the optimal subset of texture features for the learning machine (SVM).
4.1 Machine Learning
Machine learning is concerned with the development of algorithms and techniques
that readily allow computers to mimic intelligent behavior using empirical data,
such as from sensors or databases. Data from sensors is seen as different samples
that illustrate the relations between observed variables or features. A major
challenge in machine learning research is to learn how to recognize complex
patterns and make decisions based on the data. The only limitation in machine
123
learning is that, the set of all possible behaviors (patterns) given all possible
inputs, is too large to be covered by the set of trained (seen) samples (Brieman,
2001). So, the learning machine must learn (training phase) from the given
samples, so it can predict (classify) from new and unseen samples (testing phase).
The following sections discuss in detail regarding the fundamentals concepts of
machine learning and pattern recognition.
4.1.1 The Act of Learning
In humans the act of learning is namely the process of gaining knowledge or skill in
something by experience. Common and apparently simple human processes such
as: recognizing a landscape, understanding spoken words, reading handwritten
characters or identifying an object by touching it, all belie in the act of learning
(Cortes & Vapnik, 1995). In fact, the condition for a landscape to be recognized,
spoken words to be understood, handwritten characters to be read and objects to
be identified, is that the human brain has been previously trained in order to do
that, namely it has learned how to do that. This is why it is necessary to admire a
landscape several times before recognizing it from a slightly different view, or to
hear an unknown foreign word more than once before becoming familiar with it.
From the examples discussed above, it is evident that the act of learning plays a
crucial role in all the processes requiring the solution of a pattern recognition task;
all the processes in which the human brain is required to take an action based on
the class of the data it has acquired are referred to as learning. For instance,
hearing a voice and deciding whether it is a male or a female voice, reading a
handwritten character and deciding whether it is an Ä or a ℬ, touching an object
and guessing its temperature, these are typical pattern classification problems
124
(Schölkopf, 1997). These processes represent the totality of the processes a human
being has to deal with. Finding a solution for them has been crucial for humans to
survive. For that reason, highly sophisticated neural and cognitive systems have
evolved for such tasks during the last few decades. The scheme used by the human
brain to address pattern recognition tasks is based on two separate phases,
namely: a training phase and a testing phase. In the training phase the human
brain gets experience by dealing with patterns taking from the same population, as
landscapes, spoken words or handwritten characters. Then, in the test phase, it
applies to patterns of the same population, which are previously unseen. In this
sense, admiring a known landscape several times by trying to identify its
characteristics, represents the training phase, whereas recognizing it from a
slightly different view represents the test phase.
As regard to computers, the act of learning refers to artificial intelligence. In this
context, machine learning is an area of artificial intelligence that deals with the
development of techniques which allow machines to learn how to solve pattern
recognition problems; whereas learning machines are automata which solve
pattern recognition problems. In a similar way to what happens for the human
brain, the solution of a pattern recognition problem initially involves the collection
of a dataset of training patterns. The learning machine structure is then adapted
such as to create a mapping relationship between the input and output values. The
recognition performance of the trained learning machine is then evaluated on a
data of test patterns, namely patterns which were not part of the training dataset,
but were taken from the same population (Tarassenko, 1998).
125
The success of machine learning since the 1960s up to modern times is two-fold.
Firstly, it is evident that implementing learning processes by using machines is
fundamental in order to automatically address pattern recognition problems,
which due to their complexity are challenging for humans to solve. For example,
pattern classification and recognition problems such as speech recognition,
fingerprint identification, handwriting recognition, face recognition, DNA sequence
identification, video surveillance, and much more can be easily addressed by
means of learning machines. Secondly, by trying to give answers and explanations
to the numerous questions and doubts arising from the implementation of such
automatic learning systems, a deeper understanding of the processes governing
human learning is gained (Brieman, 2001).
As already discussed, all those problems requiring human or artificial intelligence
to take actions based on the data acquired are formally defined as pattern
recognition problems. This family of problems can be further divided into families
of sub-problems. The most common and important of these are pattern
classification, regression and time-series prediction problems (Tarassenko, 1998).
Pattern classification problems are those in which the learner is required to learn
how to separate the input patterns into two or more classes. A typical pattern
classification problem can require, for example, a human brain or a learning
machine to separate between two classes of patterns, such as malignant and
benign abnormalities taken from digital mammography datasets, as shown in
Figures 5.23 and 5.24 respectively. When the problem does not require associating
the class of membership to input patterns, but rather to associate a continuous
value, a regression problem is faced (Schölkopf, 1997). A typical regression
126
problem could require a human brain or machine learning algorithm to predict
prognosis as a regression of the factors associated with breast cancer using clinical
datasets. Lastly, time-series prediction problems, in which a learning machine is
trained to predict the �) + 1�cd sample in a time-series from the previous )
samples is a special case of regression problems, which assumes that the
underlying data generator is stationary, namely its statistical properties are time-
independent (Schölkopf, 1997). In this research, attention will be concentrated on
pattern classification, which is the most common type of pattern recognition
problem.
4.1.2 Learning Pattern Classification
Specific details about pattern classification, namely the way in which learning
machines address pattern classification tasks is presented in the following
sections. In particular, two important aspects of learning machines will be
discussed. First, how they learn directly from data, without using any priori
assumption on the classification problem they are facing. Secondly, how
supervised and unsupervised learning paradigms are implemented in order to
practically solve pattern classification problems.
4.1.2.1 Learning from Data
One of the most important characteristics of learning machines is that they are not
programmed by using some prior knowledge on the probability structure of the
dataset considered; in fact they are trained using repeatedly large numbers of
samples for the problem under consideration. In a sense, they learn directly from
the data how to separate the different existing classes. This approach determines
some important peculiarities of learning machines. First, they are particularly
127
suited for complex classification problems, as the whole solution is difficult to
specify a priori. Secondly, after being trained they are able to classify data
previously not encountered. This is often referred to as the generalization ability of
learning machines. Finally, since they learn directly from data, so the effective
classification solution can be constructed far more quickly than using traditional
approaches.
For this reason, an approach purely based on learning from data is regarded as the
most appropriate solution. For example, as described in Duda et al. (2001), the
Bayesian decision theory is considered a fundamental approach for solving pattern
classification problems. This theory assumes that the decision problem is given in
probabilistic terms and all the probability values are known. The Bayesian theory
quantifies the tradeoffs between different classification decisions using probability
estimates and the costs that accompany these decisions. Unfortunately, for the
most part of applications, the probabilistic structure of the problem is unknown.
Since, there is limited knowledge about the situation, the training data can be used
to classify the patterns. The approach is then to determine a way to use this
information in order to design the classifier. One approach is to use the training
patterns for estimating probabilities and densities, and use the resulting estimates
as they are the true values (Schölkopf, 1997). For this reason, with regards to the
classification of malignant and benign patterns in this research, it would be more
appropriate and effective to address the classification task using a modeling
approach rather than an approach purely based on learning from data.
128
Pattern classification in machine learning applications can be implemented using
two major schemes, namely: supervised learning and unsupervised learning. In a
supervised learning scheme, the input patterns used to train the learning machine
are labeled, in other words they are patterns whose class membership is known.
However, in an unsupervised learning scheme, the input patterns are unlabeled,
namely their class membership is unknown Duda et al. (2001).
The secondary contribution of this research outlined in Section 1.3 of this thesis, is
to demonstrate that SVMs can effectively solve pattern classification problems. The
SVM being a supervised learning scheme is discussed in detail in Section 4.2 of this
thesis. The following section provides the fundamental concepts and mathematical
details of the supervised learning scheme, which is required to fully comprehend
the methods discussed in this thesis.
4.1.2.2 Supervised Learning
Supervised learning is defined as the machine learning task of inferring a function
from supervised training data. As the training data consists of a set of training
samples, each training sample represents a pair of an input features (typically a
vector) and a desired output value or label (referred as the class membership).
Similarly, each testing sample is a pair of input features with no desired output
value (class membership) (Cortes and Vapnik, 1995).
In pattern classification problems, a supervised learning algorithm analyzes the
training data and produces an inferred function, also known as a classifier. During
the testing phase, the inferred function is used to classify the correct class
membership for any valid pair of input features. Thus, the inferred function
129
predicts the class membership of the testing samples. This requires the learning
machine to generalize using the training data to classify unseen samples in an
appropriate manner (Vapnik, 1995). In order to implement the supervised
learning scheme for a binary (two-class) classification task, a labeled dataset of
training patterns must be provided. To serve this purpose, the training patterns
(Vapnik, 1998):
�O�, … , Oh� with O\ ∈ ℝ� ∀& = 1,… , � (4.1)
are required, as well as the associated labels indicating their class membership:
�P� , … , Ph� with P\ = ±1 ∀& = 1, … , � (4.2)
Then each training pattern & is represented by a vector O\ of ) input features,
namely the individual measurable heuristic properties of the malignant and benign
patterns. For the case of this research, the heuristic properties of the patterns refer
to the classification features, namely texture features, as indicated in the proposed
framework in Figure 1.4. The classification features of each sample are associated
to a class membership of the label P\ , which takes values +1 or −1 for a binary
classification problem. For example, in a pattern classification problem in which
malignant and benign samples (see Figures 5.23 and 5.24 respectively) are
required to be separated classification features such as the: pixel intensity values
of each or some specific measurements of each image, such as: luminosity, gradient
etc. can be used. As regards to the class memberships (or labels) in this research,
all patterns belonging to the category (class) of k+�&*)+)' patterns are associated
to the label +1 and those belonging to the category of �!)&*) patterns are
associated to the label −1, as shown in Figure 4.1 (Cortes and Vapnik, 1995).
130
Figure 4.1: Supervised learning scheme
During the training phase of the supervised learning scheme, the learning machine
adjusts its internal parameters by being shown the input features O\ of the patterns
taken from k+�&*)+)' class and those patterns taken from the �!)&*) class. Once
the training phase terminates, the learning machine is supposed to have learned
how to recognize features belonging to both classes (Cortes and Vapnik, 1995). In
particular, it is supposed to have learned how to correctly separate the two
different classes in the )-dimensional feature space, as shown in Figure 4.1. At the
end, the generalization performance of the classifier is tested on a new set of
unseen data (testing data), which is not part of the training set.
4.1.2.3 Issues in Pattern Classification
In order to improve the performance of learning machines, it is most commonly
desirable to submit the data to some preprocessing techniques whose aim is to
enhance the data and reduce noises, so that the learning machine can provide
optimum classification results. Even though preprocessing techniques are not a
part of the classification task, they play a fundamental role in it. For that reason,
they are usually considered as sub-issues of pattern classification problems (Duda
et al., 2001), (Tarassenko, 1998).
131
Figure 4.2: Feature extraction. The classification problem is more easily separable
using the pair of features ��and �� (right) than using �� and �� (left)
When dealing with pattern classification problems, the first important
preprocessing step must be taken in the direction of choosing the most
appropriate features. This literally means selecting the measurable properties of
the phenomena under consideration, which can be the most discriminant features
for the specific pattern classification problem faced. For example, as indicated in
Figure 4.2, it could happen that the choice of a pair of features, say, ��and ��, makes
the classification task more easily separable than using a different pair of features,
namely �� and ��. It is evident that extracting the most discriminant feature is a
problem, which requires a priori knowledge of the data. Furthermore, the
conceptual boundary between feature computation and classification can be
considered as arbitrary. As an ideal feature extraction would yield a representation
that makes classification trivial, similarly, a powerful classifier would not need the
assistance of a feature extractor (Duda et al., 2001). Section 4.4 presents and
discusses feature selection using SVMs in detail.
The second preprocessing step consists in removing the presence of noise in the
data. The definition of noise is very general. Any property of the sensed patterns
132
due to randomness in the world or in the sensors can be considered as noise. All
non-trivial decision and pattern classification problems involve noise in some
form, since all real-world data is transduced from sensors into digital data. It is
thus unavoidable that all sensory datasets suffer from the specific noise of the
experimental set up. Typical examples are visual noise in video cameras,
background noise in audio registrations and so on.
4.1.3 Validation Techniques
The introduction of validation techniques is motivated by the willingness of finding
a solution to two fundamental problems in pattern classification. The first problem
being the selection of the learning machine’s model and the second problem being
the validation of the learning machine’s classification performance (Kohavi, 1995).
Almost invariably, all learning machines have one or more free parameters which
can be tuned in order to adapt them to each specific classification problem. For
example, in Artificial Neural Networks (ANNs) the free parameters are represented
by the number of layers in the network and by the weights linking each input
pattern to each perceptron. When a pattern classification problem is addressed by
learning machines, the typical approach consists in choosing a specific
configuration of the free parameters, namely choosing a specific model for the
learning machine and estimating its classification performance. Classification
performance is usually estimated by the so-called true error rate, literally the
learning machine’s error rate on the entire population under examination. The
configuration of free parameters for which the true error rate is minimum,
corresponds to the optimal learning machine’s model for that particular problem
(Tarassenko, 1998).
133
It is evident that in an ideal and unrealistic situation in which the dataset is
comprised of an unlimited number of patterns, the straightforward solution of the
problem will be at first to choose the learning machine’s model that provides the
lowest error rate on the entire dataset, and this error rate can be considered as the
true error rate. Obviously, in real-world applications, only finite datasets are
available and typically they are smaller than what would be desirable.
In this case, a crude approach would be to use the entire dataset to train the
learning machine in order to select the model and to estimate the error rate.
However, this crude approach suffers from two fundamental drawbacks. First, the
final model will normally overfit the training data. This means that the learning
machine results in excessive optimization on the training data, thus, it loses in
generalization performance and gives very poor interpolation on a different
dataset. Secondly, the error rate estimate is overly optimistic, typically lower than
the true error rate. It is in fact not uncommon to have 100 percent correct
classification on the training data (Duda et al., 2001). In order to overcome these
drawbacks, some sophisticated validation techniques for supervised learning
schemes have been introduced, which are presented in the following sections.
4.1.3.1 Holdout Method
An interesting approach consists of splitting the dataset into two disjoint subsets,
thus applying the holdout validation technique. The holdout method, generally
referred to as the test sample estimate, partitions the data into two mutually
exclusive subsets, known as the training set and the test set, in analogy to what has
been discussed in the previous sections. It is common to designate two-thirds (77.5
percent samples) of the dataset as the training set and the remaining one-third
134
(33.5 percent samples) as the test set, as indicated in Figure 4.3 (Kohavi, 1995).
The training set is used to train the learning machine and the trained learning
machine is then tested on the test set.
Figure 4.3: Holdout method
This method suffers from two important drawbacks. Firstly, assuming that the
learning machine’s classification performance increases as more patterns are seen,
the holdout approach is a pessimistic estimation routine because only a portion of
the data is given during the training phase. Secondly, since it is a single train and
test experiment, its estimate of the error rate can be inaccurate and misleading if it
happens to get an unfortunate split, i.e., if the test set is composed of the most
difficult patterns from the entire dataset.
4.1.3.2 Cross-validation
The �-fold Cross-Validation (CV) is a technique in which the data set É is randomly
split into � mutually exclusive folds Ê�, Ê�, … , Ê_ each having an equal size. In a
specific CV case, referred to as stratified cross-validation, the folds are stratified so
that they contain the same proportion of the samples as in the original dataset.
Here the learning machine is trained and tested � times, namely for each time
' ∈ l1,2,… , �t it is trained on É\Éc and tested on Éc, as shown in Figure 4.4
(Kohavi, 1995).
135
The major advantage of the CV technique with respect to the holdout method is
that all the patterns in the dataset are used for training and testing. At the same
time, the true error is estimated as the average error on the test patterns, thus
preventing the problems arising from unfortunate splits. The average error rate for
�-fold CV can be calculated using the following expression (Kohavi, 1995):
!0Ì� = �_∑ !\_\s� (4.3)
where � represents the number of folds and !\ represents the true error rate for
each of the �-folds.
Figure 4.4: �-fold cross-validation method
Some interesting considerations concerning the choice of the correct number of
folds � are drawn in Kohavi (1995). First, if the number of folds � is large, the bias
of the true error estimate is generally small, thus the estimator may be considered
very accurate. Unfortunately, due to the large number of iterations, the variance of
the true rate estimator as well as the computational time is expected to be large.
136
Secondly, if the number of folds � is reduced, the bias of the true error estimate is
generally large, thus the estimator may be considered conservative or higher than
the true error rate. In such cases, due to the reduced number of iterations, the
variance of the true error rate estimator as well as the computational time is
typically small. In practice, the choice of the number of CV folds strongly depends
on the size of the dataset. For large datasets, even a 3-fold CV could be quite
accurate. For sparse datasets, it may be necessary to partition the dataset in a large
number of folds to train on as many patterns as possible. A common choice for �-
fold cross validation is typically, � = 10.
Figure 4.5: Leave-one-out method
4.1.3.3 Leave-One-Out Method
For very sparse datasets, the leave-one-out validation method is simple to
implement. This approach is analogous to that of CV. The only difference is that, if
U is the number of patterns in the dataset, the learning machine is trained U times.
In particular, for each time, U − 1 patterns are used for training and the remaining
137
for testing, as shown in Figure 4.5. Similar to CV, the true error rate in leave-one-
out validation is estimated as the average error rate on the test patterns Kohavi
(1995).
4.2 Support Vector Machine
The Support Vector Machine (SVM) is one of the advanced techniques amongst the
many learning algorithms deeply inspired by the statistical learning theory, which
appeared in the machine learning community in the last decade. Its original
formulation is quite recent and is mainly due to Vapnik & Chervonenkis (1974),
Boser et al. (1992), Guyon et al. (1993), Cortes & Vapnik (1995), Vapnik (1995,
1998).
During the 1990s many machine learning algorithms arose contradicting the
biological paradigm, since most were inspired by the minimization of theoretical
bounds on the error rate. The SVM is not an exception to that. In fact, for a given
learning task and with a finite amount of training patterns, the SVM is a learning
machine which achieves its best generalization performance by finding the right
balance between the accuracy obtained on that particular training set and the
complexity of the machine, namely its ability in learning any training set without
errors (Burges, 1988).
The following sections present introductory notions on the statistical learning
theory and the mathematical concepts of SVMs, in order to demonstrate that
finding the right balance between accuracy and capacity is equivalent to finding
the minimum of a theoretical bound on the error rate.
138
4.2.1 Statistical Learning Theory
Suppose that, � training patterns are represented by (Vapnik, 1998):
O = �O�, . . . , Oh� with O\ℝ� ∀& = 1,… , � (4.4)
together with the associated class labels representing the attended class
membership:
P = �P�, . . . , Ph� with P\ = ±1 ∀& = 1,… , � (4.5)
Assume that patterns are generated independently and distributed identically
according to an unknown probability distribution ��O, P�. As already known,
learning machines address the task of pattern classification by finding a rule which
assigns each input pattern to a class membership. In particular, during the training
phase, a mapping �:ℝ� → l±1t is created between input patterns and labels, such
that the learning machine is expected to correctly classify unseen test examples.
The best mapping of � that can be obtained is by minimizing the expected error,
represented by the following expression (Vapnik, 1998):
�� = Î �� |P − ��O�| ���O, P� (4.6)
where the component |P − ��O�|���O, P� is known as the loss function. A further
common loss function for example is �P − ��O���, which is also known as the
squared loss function.
139
Figure 4.6: Overfitting phenomenon. The more complex function obtains a smaller
training error than the linear function (left). But only with a larger data set it is
possible to decide whether the more complex function really performs better
(middle) or overfits (right)
Unfortunately, the expected error cannot be directly minimized, since the
probability distribution function ��O, P�, from which the data is generated, is
unknown. In order to estimate a function � that is close to the optimal one, an
induction principle for risk minimization is therefore necessary. The most
straightforward way is to approximate the minimum of the risk discussed in
equation (4.7) by the minimum of the so called empirical risk, namely the
measured mean error rate on the training set (Vapnik, 1998):
�� �� = �h ∑ �� |P − ��O\�|h\s� (4.7)
A small error on the training set does not necessarily indicate a high generalization
ability. As anticipated in Section 4.1.3, this phenomenon is known as overfitting, as
shown in Figure 4.6. In particular, as described in Müller et al. (2001), given a
small training dataset as shown in the left of Figure 4.6, functions of � with higher
degrees of complexity may result in smaller training errors. Nevertheless, only
with a larger dataset, as the ones shown in the middle and right of Figure 4.6, it is
understood that the decision on the right of Figure 4.6, reflects the true
distribution more closely and does not overfit. One technique to avoid overfitting,
is by restricting the complexity of the function �. This means that, a simple linear
140
function, describing most of the data is preferable to a complex function. These
considerations, give rise to the problem of how to determine the optimal
complexity of a function (Vapnik, 1995).
A specific technique introduced by Vapnik for controlling the complexity of a
decision function is represented by the Structural Risk Minimization (SRM)
principle (Vapnik, 1979). In order to understand how SRM works, it is necessary to
introduce a non-negative integer a referred to as the Vapnik-Chervonenkis
dimension or VC-dimension, which describes the complexity of a class of functions.
In particular, the SRM measures the number of training points that can be
separated for all possible labels using functions of that class. Once the concept of
the VC-dimension is introduced, a nested family of function classes must be
constructed (Vapnik, 1998):
F� ⊂ F� ⊂… ⊂ F_ (4.8)
whose VC-dimension satisfies:
a� ≤ a� ≤…≤ a_ (4.9)
Then, suppose that the solutions of the empirical risk minimization problem:
�� ≤ �� ≤…≤ �_ (4.10)
respectively belong to the function classes, Ê\, & = 1,… �. In that context, the SRM
principle chooses the function �\ in the class Ê\ such that the right-hand side of the
following bound on the generalization error is minimized (Vapnik, 1998):
141
�� ≤ �� �� + ÑÒd�=5L²ÓÔ2��/=5L�ÕÖ�h × (4.11)
where a is the VC-dimension of the function class under consideration, the square
root term is called the confidence term and the bound holds with probability 1 − Ø
for any 0 ≤ Ø ≤ 1.
Figure 4.7: Schematic illustration of the bound in equation (4.11). The dotted line
represents the empirical risk �� ��. The dashed line represents the confidence
term. The condintuous line represents the expected risk ��. The best solution is
found by choosing the optimal tradeoff between the confidence term and the
empirical risk �� ��
Three aspects are of great interest in the above bound in equation (4.11). First, it is
independent of ��O, P�, as it assumes only that the entire dataset (training and test
set) is drawn independently according to some ��O, P�. Secondly, it is usually not
possible to compute the left-hand side, namely the expected risk. Thirdly, the
known VC-dimension a on the right-hand side is easily computable. This means
that the selection of the learning machine which maps the input patterns to their
class memberships by the function �:ℝ� → l±1t and minimizes the right-hand
142
side of equation (4.11), corresponds to the selection of the learning machine that
achieves the lowest upper bound on the expected risk (Vapnik, 1995).
Minimization of the expected risk �� is generally achieved by obtaining a small
training error �� �� while keeping the function class to be small as possible.
However, two extreme situations may arise. Firstly, a very small function class
gives a disappearing square root term, but a relatively large training error.
Secondly, a huge function class gives a disappearing empirical error, but a larger
square root term. Nevertheless, from these considerations, it is evident that the
best solution of the problem is usually in between, as shown in Figure 4.7. In other
words, finding the minimum of the expected error actually means finding the right
tradeoff between the accuracy obtained on that particular training set and
complexity of the mapping created by the learning machine(Vapnik, 1995).
In practical problems the bound on the expected error discussed in equation (4.11)
is neither easily comparable nor very helpful. A typical problem is that the VC-
dimension of the class under consideration is either unknown or infinite. In such a
case an infinite number of training data would be necessary. Nevertheless, the
existence of bounds is important from a theoretical point of view, since it offers
some deeper insights into the nature of learning machines.
4.2.2 Linking Statistical Theory to SVM
Linear learning machines such as the perceptron and SVM, as it will be clarified in
the next section, use hyperplanes to separate classes in the feature space as shown
in Figure 4.8, and can be represented mathematically as a function in the form
(Vapnik, 1979):
143
P = sign�Ù ∙ O + �� (4.12)
In Vapnik and Chervonenkis (1974) it has been demonstrated that the VC-
dimension can be bounded in terms of another quantity referred to as a margin,
namely the minimal distance of patterns from the hyperplane.
Figure 4.8: A hyperplane separating different patterns. The margin is the minimal
distance between the pattern and the hyperplane, thus here the dashed lines
In Figure 4.8 the margin corresponds to the dashed lines. In particular, by rescaling
Ù and � such that the points that are closest to the hyperplane satisfy the condition
|Ù ∙ O + �| = 1, namely transforming the hyperplane to its canonical
representation, it is possible to measure directly the margin as a function of Ù.
Consider, two patterns O� and O� belonging to two different classes and such that
|Ù ∙ O� + �| = +1 and |Ù ∙ O� + �| = −1. Then the margin can be calculated as the
distance between those two points along the perpendicular, namely as Vapnik and
Chervonenkis (1974):
144
Û‖Û‖ �O� − O�� = �‖Û‖ (4.13)
Then, it can be demonstrated that the inequality which links the VC-dimension of
the class of separating hyperplanes to the margin uses the following expression:
a ≤ Λ�� + 1 and ‖Ù‖ < Λ (4.14)
where represents the radius of the smallest ball around the data. Due to the
inverse proportionality between the margin and ‖Ù‖ (in equation (4.13)), a small
VC-dimension is obtained by requiring a large margin. On the other hand, a high
VC-dimension is obtained by requiring a small margin. The bound described in
equation (4.11) demonstrates that in order to achieve a small expected error it is
necessary to keep both the training error and the VC-dimension small, then when
working with linear learning machines, separating hyperplanes can be constructed
such that they maximize the margin and separate the training patterns with as few
errors as possible (Vapnik and Chervonenkis, 1974). As it will be demonstrated in
the following section, this result forms the basis of the SVM learning algorithm.
4.2.3 Linear SVM
4.2.3.1 Separable Case
In order to introduce the SVM learning algorithm, the simplest case to deal with, is
the separable case in which data is linearly separable. As it will be discussed in the
following section, the most general case, namely, the non-linear SVM trained on
non-separable data results in a similar solution. Suppose again that � training
patterns are given (Vapnik and Chervonenkis, 1974):
145
�O�,… , Oh� with O\ ∈ ℝ� ∀& = 1,… , � (4.15)
together with the associated class labels representing the attended class
membership:
�P� , … , Ph� with P\ = ±1 ∀& = 1, … , � (4.16)
Assume that the patterns are linearly separable, namely they could be separated
by a hyperplane P = sign�Ù ∙ O + �� as the one shown in Figure 4.8. For such a
learning machine, the conditions for classification without the training error rate
are:
P\�Ù\ ∙ O\ + �� ≥ 1 & = 1,2,… , � (4.17)
The aim of learning is thus finding Ù and � such that the expected risk is
minimized. According to equation (4.11), one strategy that can be implemented, is
to make the empirical risk zero by forcing Ù and � to result in a perfect separation
between the two classes, while at the same time minimizing the complexity term of
the VC-dimension a. For the case of a linear learning machine the VC-dimension a
is bounded as described in equation (4.14), so it is possible to minimize the VC-
dimension by minimizing the ‖Ù‖� term, namely by maximizing the margin
(Vapnik and Chervonenkis, 1974). The linear learning machine ensures the lowest
expected risk is that, which gives an empirical risk zero, or perfect separation
between the two classes represented by equation (4.17) and at the same time
minimizes the VC-dimension or maximizes the margin between the two classes
using the expression:
146
minÛ,à �� ‖Ù‖� (4.18)
In order to solve this convex quadratic programming (QP) optimization problem, it
is preferable to introduce a Lagrangian ℒ (Vapnik and Chervonenkis, 1974):
ℒ�Ù, �, â� = �� ‖Ù‖� −∑ â\�P\�Ù\ ∙ O\ + �� − 1�h\s� (4.19)
with the Lagrangian multipliers satisfying â\ > 0, & = 1,2,… �. The Lagrangian ℒ has
to be minimized with respect to Ù and � and to be maximized with respect to â\ . The conditions that at the saddle points the derivatives vanish are as follows:
äℒ�Û,à,å�äà = 0 (4.20)
äℒ�Û,à,å�äÛ = 0 (4.21)
leads to:
∑ â\P\h\s� = 0 (4.22)
Ù = ∑ â\P\h\s� O\ (4.23)
By substituting equation (4.23) in equation (4.19), the dual quadratic optimization
problem is obtained:
maxå ∑ â\ − ��h\s� â\âªP\Pª�O\ ∙ Oª� (4.24)
subject to the constraints:
â\ > 0,& = 1,2,… , � (4.25)
∑ â\P\ = 0h\s� (4.26)
147
Thus, solving the dual optimization problem, the Lagrangian multipliers â\ > 0,
& = 1,2,… �, need to express the specific Ù which solves equation (4.18). In
particular, for each input pattern O, the following decision function will be applied
(Vapnik, 1998):
��O� = sign�∑ â\P\h\s� �O ∙ O\� + �� (4.27)
The hyperplane given by the decision function in equation (4.27) is generally
referred to as the maximal margin hyperplane of linear SVM.
4.2.3.2 Non-separable Case
When dealing with noisy data, it could happen that the patterns are not linearly
separable. In such a situation it is impossible to keep the empirical error zero,
therefore, it is necessary to find the best tradeoff between the empirical risk and
the complexity term as discussed in equation (4.11). In order to relax hard-margin
constraints, thus allowing classification errors, slack variables must be introduced
(Cortes & Vapnik, 1995) such that:
P\�Ù\ ∙ O\ + �� ≥ 1 − æ\ , æ\ > 0 & = 1,2,… , � (4.28)
In this case, the solution is found by minimizing the VC-dimension and an upper
bound on the empirical risk, which represents the number of training errors. Thus
the quantity to minimize is:
minÛ,à,ç �� ‖Ù‖� + � ∑ æ\h\s� (4.29)
148
In analogy to equation (4.18), finding a minimum of the first terms actually means
minimizing the VC-dimension of the class of functions under consideration. On the
other hand, the term ∑ æ\h\s� is an upper bound, representing the number of
misclassifications on the training set, so finding a minimum for it results in
minimizing the empirical risk. In this context, the SVM regularization parameter
� > 0 determines the tradeoff between the complexity term and the empirical risk.
As in the linearly separable case, the Lagrangian multipliers for the non-separable
case are obtained by solving the following quadratic problem (QP) (Vapnik and
Chervonenkis, 1974):
maxå ℒ ∑ â\ − ��h\s� ∑ â\âªP\Pª�O\ ∙ Oª�h\,_s� (4.30)
subject to the constraints:
0 ≤ â\ ≤ �,& = 1,2,… , � (4.31)
∑ â\P\ = 0h\s� (4.32)
Thus the only difference of the non-separable case from the separable case is that
the Lagrangian multipliers are upper bounded by the constant term, � > 0. The
Karush-Kuhn-Tucker (KKT) conditions (Lasdon, 1970), state the necessary
requirements for a set of variables to be optimal for an optimization problem, only
those Lagrangian multipliers â\ , & = 1,2,… , � corresponding to a training pattern O\ which are either on the margin or inside the margin area are considered as non-
zero. The KKT conditions assert that (Vapnik, 1995):
â\ = 0, ⇒ P\��O\� ≥ 1 and æ\ = 0 (4.33)
0 < â\ < �, ⇒ P\��O\� = 1 and æ\ = 0 (4.34)
â\ = �, ⇒ P\��O\� ≤ 1 and æ\ ≥ 0 (4.35)
149
These considerations reveal a fundamental property of SVM, namely that the
solution found is sparse in â. This is crucial for computational time, since sparsity
guarantees that the expansion discussed in equation (4.23) is calculated on the
restricted number of patterns O\ corresponding to â\ > 0, also known as Support
Vectors (SVs) (Vapnik, 1995). The KKT conditions are also useful to compute the
threshold � in equation (4.27). From equation (4.34), it follows that (Vapnik,
1995):
P\�∑ âªPªhªs� �O\ ∙ Oª� + �� = 1 (4.36)
4.2.4 Non-linear SVM
SVM can afford to use more complex decision functions by remapping the input
patterns into a higher dimensional space, in which the classification between the
two classes can be performed by a separating hyperplane (Vapnik, 1998):
Φ:ℝ� → ℋ (4.37)
x → Φ�O� (4.38)
Suppose for example that some non-linearly separable patterns are given in two
dimensions, as shown in Figure 4.9. By remapping them into a three-dimensional
space of the second-order monomials (Vapnik, 1998):
Φ:ℝ� → ℝ� (4.39)
�O�, O�� → ��O���, √2O�O�, �O���� = �~�, ~�, ~�� (4.40)
the result is a linear hyperplane separating those patterns, as shown on the right
side of Figure 4.9.
150
Figure 4.9: Non-linearly separable patterns in two-dimensions (left). By
remapping them in a three dimensional space of the second order monomials
(right) a linear hyperplane separating those patterns can be found
The SVM optimization problem generally involves the dot products of the training
patterns O\, as it is evident from equation (4.27). Therefore, the non-linear
mapping Φ:ℝ� → ℋ that maps the patterns O\ into the new space known as a
Hilbert space ℋ, does not require to be given explicitly. It is however necessary to
specify the dot product of any of the two images Φ�O� and Φ�P� in ℋ space
through a ì!b)!��í)î'&$)�ì� defined over � × �, where � is a compact subset of
ℝ� including the training and test patterns. The kernel function ì can then be
defined by the following expression (Vapnik, 1998):
ì�O, P� ≡ Φ�O� ∙ Φ�P� (4.41)
In order to assure that the kernel function definition in equation (4.41) is well
posed, ì�O, P� must satisfy Mercer’s conditions (Mercer, 1909). More specifically,
ì�O, P� must be symmetric and continuous over � × �. Once Mercer’s conditions
are satisfied, it is relatively easy to find a mapping Φ of the input patterns onto the
Hilbert space ℋ (Mercer, 1909). The maximal margin hyperplane can then be
151
represented in terms of the input patterns in ℝ� , resulting in the following
expression for the decision function (Vapnik, 1998):
��O� = sign�∑ â\P\h\s� ì�O, O\� + �� (4.42)
Many kernel functions have been developed that can be used with SVMs. The
kernels summarized in Table 4.1 (in equations (4.43) to (4.45)) satisfy Mercer’s
conditions (Mercer, 1909) and are commonly used with SVM. The sigmoidal kernel
(in equation (4.43)) will only satisfy Mercer’s condition for particular values of the
free parameters (Burges, 1988), (Smola & Schölkopf, 2004), but this kernel has
been used successfully in practice (Vapnik, 1995). The polynomial kernel of degree
p (in equation (4.44)), is inhomogeneous such that. it allows the additive constant
� to be larger than zero (Burges, 1988) for additional degrees of freedom (Boset et
al., 1992).
Table 4.1: Non-linear kernels commonly used to perform a dot product in a mapped feature space in the SVM formulation
Name Parameters Kernel Function Equation
No.
Polynomial î ∈ ℝ, # ∈ ℕ ì�ñò, ñó� = �ñò ∙ ñó + î�� (4.43)
Radial Basis Function (RBF) Gaussian kernel
� ∈ ℝ ì�ñò, ñó� = !/ôõñò/ñóõ² (4.44)
Sigmoidal ö ∈ ℝ, u ∈ ℝ ì�ñò, ñó� = tanh�ö�ñò ∙ ñó� − u� (4.45)
The Radial Basis Function (RBF) kernel (in equation (4.45)) also known as the
Gaussian kernel is the most widely used kernel with SVMs. The RBF kernel is
translation invariant, this means that, ìô�ñò, ñó� = ìô�ñò − ñó� has an infinite
152
number of dimensions (Vapnik, 1995), (Burges, 1988). A significant advantage of
the RBF kernel is that it adds only a single free parameter � > 0, which controls
the width of the RBF kernel as � = 1 2±�⁄ , where ±� is the variance of the
resulting Gaussian hypersphere. The RBF kernel has been shown to perform well
in a wide variety of practical applications, such as in Degroeve et al. (2005), Hsu et
al. (2003) and Wang et al. (2005).
Using the RBF kernel with SVM, there are two SVM hyperplane parameters that
need to be determined in the SVM model which are: � and �. In order to get good
generalization ability, a validation process must be performed in order to decide
the optimum values of these parameters. The procedure for SVM hyperparameter
optimization presented by Hsu et al. (2003), which is known as the Grid Search
method is as follows:
1. Consider a grid space of ��, �� with log�� ∈ l−1,−2,… ,20t and
log�� ∈ l−20,−19,… ,1t. 2. For each SVM hyperparameter pair ��, �� in the search space, perform 10-
fold CV on the training set.
3. Choose the parameters ��, �� that lead to the highest (optimum) CV
accuracy and lowest error, and use them to build the SVM classification
engine (trained model).
4.2.5 Implementation of SVM
As discussed in Sections 4.2.3 and 4.2.4 previously, the training of SVM requires
the solution of a convex quadratic programming (QP) optimization problem. This
task is very difficult to implement by average engineers and the training
153
algorithms that use numerical QP are slow, especially for large size problems. In
the past decade, a number of researchers introduced new learning algorithms that
use faster and simpler-to-implement methods for solving QP problems in SVMs.
These methods include the: Newton method, Quasi Newton method, Kernel
Adatron (KA) (Frieß et al., 1998) and Sequential Minimal Optimization (SMO)
(Platt, 1998 & 1999a).
This research focuses on SMO because other optimization methods have proven to
be slower (Platt, 1999b) on large of datasets and also since the LIBSVM (Chang &
Lin, 2010) library used for implementing SVM in this research also uses SMO. The
following section gives a brief description of the SMO algorithm and its
implementation for SVM.
4.2.5.1 Sequential Minimal Optimization
As the training of SVMs under normal circumstances requires a lot of time for
calculation of the Kernel matrix, the training time increases tremendously when a
large number of training samples are present, resulting in a bigger Kernel matrix.
In order to solve this problem, SMO deals with the large QP problems, by breaking
(decomposing) the problem into a series of smaller QP problems.
The Sequential Minimal Optimization (SMO) algorithm was first introduced by
John C. Platt in 1998 (Platt, 1998 & 1999a). The main concept of SMO is to break
large QP problems into smaller QP problems. More specifically, in SMO a minimal
subset of only two training samples can be optimized on each training iteration.
This is because the smallest number of Lagrangian multipliers that can be used for
optimization at each step, is two. Thus, in this way each small QP problem is solved
154
analytically without the need of performing time consuming numerical
optimization, which makes implementation of SMO easy and simple.
Table 4.2: Summarized procedure of the SMO algorithm
Step
No. Procedure
1 Choose the first Lagrange multiplier to be a KKT violator.
2 Choose the second Lagrange multiplier using heuristics.
3 Update the second Lagrange multiplier via: â��Û = â� + z²�øù/ø²�ú
4 Clip the multiplier â��Û to â��Û,ûh\��ü
5 If the multiplier does not change, go back to Step 1.
6 Update the first Lagrange multiplier.
7 Update the error-cache.
8 If all Lagrange multiplier fulfill KKT conditions, stop; else go to Step 1.
At every step of SMO, two Lagrange multipliers are selected for optimization and
after their optimal values are found given that all the other multipliers are fixed,
the SVM is updated accordingly. In SMO, the two training samples are selected
using a heuristic method, and then the two Lagrange multipliers are solved
analytically. The SMO algorithm is summarized and presented in Table 4.2 (Platt,
1998 & 1999a).
The main advantage of SMO is that it uses only two training samples at every step
and avoids the computation of a Kernel matrix. Due to this reason, SMO requires a
smaller amount of memory and can handle very large training sets compared to
other optimization techniques (Platt, 1998 & 1999a) such as the chunking
algorithm (Vapnik, 1979).
155
4.3 Artificial Neural Networks
In order to estimate the performance of the SVM and perform a comparative
research as outlined in the fifth research objective in Section 1.2, different machine
learning algorithms other than SVM are evaluated in this research. Since ANNs
have similar structure to that of SVMs, so they are used for the purpose of model
comparison.
In this research two ANN based approaches are identified, i.e. a traditional and a
modern ANN approach, namely the Back-Propagation Neural Network (BPNN)
(Rumelhart et al., 1986) and the Extreme Learning Machine (ELM) (Huang et al.,
2006a) respectively. The Online-Sequential Extreme Learning Machine (OS-ELM)
(Liang et al., 2006) is a recently proposed variant of the ELM (Huang et al., 2006a),
which overcomes the limitations of standard ELM. The BPNN (Rumelhart et al.,
1986) and the OS-ELM (Liang et al., 2006) as discussed in the following sections
are adopted in this research.
4.3.1 Back-Propagation Neural Network (BPNN)
Back-propagation (BP) also referred to as “propagation of error” is a common
method used for teaching ANNs on how to perform given tasks. The BP algorithm
was first presented by Paul Werbos in 1974, however, it wasn't until 1986 that it
gained importance in the field of machine learning research (Rumelhart et al.,
1986). The BP algorithm is most suitable for use with feed-forward networks.
Feed-forward networks are those type of networks that have no feedback.
The BP is a supervised learning method, which is an implementation of the Delta
rule. The delta rule in BP typically requires a teacher (or trainer) that knows, or
156
can calculate, the desired output for any possible input. In the BP algorithm, the
errors propagate backwards from the output nodes of the network to the inner
nodes of the network. So the BP algorithm is generally used to calculate the
gradient error of the network. The gradient of the error is typically used in a
gradient descent algorithm to determine the weights that minimize the training
error (Rumelhart et al., 1986). BP networks are typically multi-layer ANNs, usually
with an input layer, one or more hidden layers and an output layer. For the hidden
layer neurons to serve any useful purpose, they must have non-linear activation
(or transfer) functions. The most common non-linear activation functions include
the: log-sigmoid, tan-sigmoid, Gaussian and softmax transfer functions.
Figure 4.10: General architecture of a Back-Propagation Neural Network (BPNN)
In a Back-propagation Neural Network (BPNN), learning is formulated as follows.
Firstly, a training pattern is presented to the input layer of the BPNN. The network
1
Hidden
Layer
. . .
. . .
. . .
. . .
. . .
. . .
2
i
n
1
2
j
m
1
2
k
l
Input
Layer
Output
Layer
y1
y2
yk
yl
x1
x2
xi
xn
Input Signals
Error Signals
wij wjk
157
propagates the input pattern from layer to layer until the output pattern is
generated by the neurons in the output layer. If the output pattern is different from
the desired output, an error is calculated. This error is then propagated backwards
through the network to the input layer. As the error is propagated backwards, the
weights connecting the neurons are adjusted by the BP algorithm. Figure 4.10
shows a typical architecture of a BPNN (Wen et al., 2000). The following steps
illustrate the implementation of the BP training algorithm.
Step 1 ― Network Initialization
All the weights and threshold levels of the network are set to small random
numbers uniformly.
Step 2 ― Activation Function
The BPNN is activated by applying inputs O����, O����,… , O���� and desired
outputs P����, P����, … , P���� as shown in Figure 4.10. Then the actual output
of the neurons in the hidden layer is calculated using the following transfer
function:
Pª��� = �$*%&*ý∑ O\��� ∙ Ù\ª��� − ¥ª�\s� þ (4.46)
where ) is the number of inputs of neuron j in the hidden layer and �$*%&*
represents the log-sigmoid activation function.
Step 3 ― Weight Adjustment
The weights in the BPNN are updated and the errors associated with output
neurons are propagated backward.
158
Step 4 ― Iteration
The value of k is increased by one, and Step 2 is repeated again. The iterations
continue until the error becomes zero or a desired performance goal is met,
whichever comes first.
As the BP algorithm is a supervised machine learning algorithm, it is applicable for
the task of classification and regression. Many complex activation functions have
been proposed by researchers in the past few decades to overcome the limitation
of standard activation functions, and the BP algorithm is applicable to all. For
further reading on the BP training algorithm see Rumelhart et al. (1986).
4.3.2 Online-Sequential Extreme Learning Machine (ELM)
The Extreme Learning Machine (ELM) was proposed by Huang in 2006 (Huang et
al., 2006a) for Single Layer Feed-forward Networks (SLFNs), which states to
produce superior performance (Huang and Chee-Kheong, 2004), (Huang et al.,
2006b), (Huang et al., 2006c) compared to other machine learning algorithms. The
ELM is one algorithm amongst the supervised batch learning algorithms that uses
a finite number of input and output samples for training. The ELM algorithm is
claimed to be extremely fast in its learning speed and has better generalization
performance compared to conventional learning algorithms (Huang et al., 2004).
The ELM is a modern learning algorithm for SLFNs that works efficiently for
classifications, function approximations and online prediction problems. Moreover,
the ELM can work well for a variety of types of applications. Typically a SLFN has
three input parameters: (i) input weight Ù\ , (ii) hidden neuron biases �\ , and (iii)
output weight �\ . While traditional learning algorithms of SLFNs have to tune these
159
parameters, the ELM randomly generates the input weight Ù\ , the hidden neuron
biases �\ and then calculates the output weight �\ . Thus, for SLFNs trained using
the ELM algorithm, no further learning is required.
Given U arbitrary distinct samples �O\ , '\�, where O\ = O\�, O\�, … , O\��� ∈ ℝ� and
'\ = '\�, '\�, … , '\��� ∈ ℝ� , a SLFN with U� hidden neurons and transfer function
*�O� can be represented by:
∑ �\*�Ù\ ∙«�\s� Oª + �\� = $ª,j = 1,2,… ,U (4.47)
where Ù\ represents the weight vector connecting the input neurons and the &7�
hidden neuron i, �\ is the threshold of the &7� hidden neuron and �\ is the weight
vector connecting the &7� hidden neuron and the output neurons. In equation
(4.47) Ù\ ∙ Oª represents the inner product of Ù\ and Oª . If the network is able to
approximate these U samples with zero error, then ∑ õ$ª − 'ªõ«ªs� = 0,, there exists
�\ , Ù\ , �\ such that, ∑ �\*�Ù\ ∙«�\s� Oª + �\� = 'ªwherej = 1,2,… , U. Thus, the above
equations can be represented as ¹� = �, where:
¹�Ù�, … , Ù«� , �� , … , �«� , O�, … , O«�� =
�*�Ù� ∙ O� + ��� ⋯ *�Ù«� ∙ O� + �«��⋮ ⋯ ⋮*�Ù� ∙ O« + ��� ⋯ *�Ù«� ∙ O« + �«���«×«� (4.48)
� = ����⋮�«���«×� and � = �'��⋮'«���«×� (4.49)
160
As discussed in Huang and Babri (1998), H is the hidden layer output matrix of the
ANN, with the &7� column of H being the &7� hidden neuron output for the inputs
O�, O�, … , O�. Based on the previous work of Huang, matrix H is considered to be
square and invertible only if the number of hidden neurons equals the number of
distinct training samples U� = U. Satisfying the above condition for matrix H
indicates that SLFN can approximate the training samples with almost a zero error.
Typically the number of hidden neurons is much lower than the number of distinct
training samples, U� ≪ U. In equation (4.48) H is a non-square matrix, which
indicates that there may not exist Ù\ , �\�\�& = 1,… , U� such that ¹� = �. Thus, a
specific set of Ù\ , ��\��\�& = 1,… , U�� needs to be found so that:
õ¹�Ù� ,… , Ù«� , ��� , … , ��«���� − �õ
= minÛ|,à|,�|‖¹�Ù�, … ,Ù«� , ��, … , �«��� − �‖ (4.50)
which is equivalent to minimizing the cost function,
= ∑ �∑ �\*�Ù\ ∙«�\s� Oª + �\� − 'ª��«ªs� (4.51)
Huang in (Huang et al., 2006b) and (Huang, 2003) discussed that the hidden
neuron parameters in the ELM need not be tuned, as the matrix H converts the
data from non-linear separable cases to high-dimensional linear separable cases.
Furthermore, Huang in (Huang et al., 2004) showed that the input weights and
hidden neurons need not to be tuned and can be randomly selected and then fixed.
Thus, for fixed input weights and the hidden layer biases (kernel parameters),
training a SLFN is equivalent to finding a least squares solution �� for the linear
system, ¹� = �.
161
In order to handle online applications a variant of the ELM typically known as the
Online-Sequential Extreme Learning Machine (OS-ELM) was introduced by Liang
et al. (2006). The OS-ELM was proposed to overcome the limitations of ELM as
developed by Huang (Huang et al., 2004). Since the ELM algorithm belongs to
supervised batch learning algorithms, this prohibits its further application. As in
the real world, training data may arrive either chunk-by-chunk or one-by-one,
therefore, an online-sequential learning is most suitable to cater for such
variations. The OS-ELM was originally developed for SLFNs with additive or RBF
hidden nodes in a unified framework, thus it can handle both additive neurons and
Radial Basis Function (RBF) nodes. Unlike other sequential learning algorithms
that require many parameters to be tuned, the OS-ELM only requires the number
of hidden layer neurons to be specified. Huang et al. (2004) proposed that in order
to determine an optimum value for the parameter ), an iterative approach needs
to be applied, which is as follows:
Start with an initial value of ) = 20
Increment ) by 20 on each iteration.
Calculate the training accuracy of the model using for ).
Stop iterations when ) = 200
The OS-ELM algorithm as proposed by Liang et al. (2006) consists of two major
phases, namely the, initialization (or boosting) phase and the sequential-learning
phase. In the initialization phase, the number of data required should be equal to
the number of hidden layer neurons. The initialization phase trains the SLFN using
the OS-ELM method given by some batch of training data. This data is discarded
once the process is complete. Following the initialization phase, in the learning
162
phase the OS-ELM learns the training data using a chunk-by-chunk procedure. All
the training data is discarded once the learning procedure involving the data is
complete.
The OS-ELM batch training algorithm, provides a faster learning capability as
compared with traditional machine learning techniques. Unlike other popular
learning machines, only a small amount of human involvement is required in
implementing the OS-ELM. Except for the number of the hidden neurons
(insensitive to OS-ELM), no other network parameters need to be optimized by the
users, since the OS-ELM algorithm chooses the input weights randomly and
analytically determines the output weights itself Huang et al. (2006b).
4.4 Recursive Feature Elimination
A challenging task in pattern classification problems, namely machine learning, is
to reduce the dimensionality ) of the feature space by finding a restricted number
of features yielding good classification performance. In recent years, a lot of work
has been done in the direction of feature selection (Kohavi and John, 1997),
(Kearns et al., 1997). Feature elimination has known to be a fundamental process
in order to reduce the computation time required to solve pattern classification
problems and to improve the classification performance of the learning machine.
The curse of dimensionality from statistical theory point of view, asserts that the
difficulty of a prediction problem increases with the dimension ) of the feature
space, as in principle, exponentially many patterns are required to sample the
space properly.
163
The SVM is introduced in this chapter as a tool for the solution of pattern
classification problems. Next, new aspects of the applicability of SVMs in
knowledge discovery and data mining, namely feature selection will be discussed.
The reason for this is that SVMs are known to be very effective for discovering
informative attributes of the dataset, namely critically important features. To serve
this purpose, feature selection methods, namely, Recursive Feature Elimination
(RFE) will be discussed.
Since, feature ranking is the first task addressed towards the elimination of
unimportant features, this forms the basis for using Non-linear SVM (in Section
4.2.4) for the purpose of pattern classification. The algorithms presented in the
following sections are usually known for combining SVM-based RFE using various
strategies such as the F-score and the Random Forest (RF) (Chen and Lin, 2006).
4.4.1 Feature Ranking Using F-score
Feature ranking methods define the importance of each single feature according to
its contribution to the learning machines’ predictive accuracy. The final aim is to
obtain a ranked list of features from which the features having an important
contribution to the model can be selected, whereas features having a smaller
contribution can be eliminated. Thus, feature ranking eliminates all those features
which are useless for discrimination purposes, or at least represent noise.
Several methods evaluating how well individual features contribute towards a
binary classification (two-class) problem have been indicated in the literature. For
instance, Golub et al. (1999) used the following correlation coefficient �b\� as the
feature ranking criterion:
164
b\ = ®|����/®|����¯|����2¯|���� (4.52)
where °\ and ± \ are respectively the mean and the standard deviation of the
feature & for all the patterns whose class is positive �#0� or negative �)0�. Large
positive b\ values represent a strong correlation with the class #0, whereas large
negative b\ values represent strong correlation with class )0 . Then, by selecting an
equal number of features with positive and negative correlation coefficients, the
two classes can be represented. Another approach, as described by Furey et al.
(2000) used the absolute value �b\�, whereas the authors in Pavlidis et al. (2001)
used the following correlation coefficient:
b\ = �®|����/®|�����²�¯|����2¯|�����² (4.53)
An important drawback characterizing these feature ranking techniques (in
equations (4.52) and (4.53)) is that, they rely on the implicit orthogonality
assumptions that they make. In fact, each correlation coefficient �b\� is computed
by using only the information on that single feature, thus without taking into
account the mutual information between features. This is a major problem, since
features are typically correlated with each other, as the case with pattern
classification problems. In order to overcome this problem, it is necessary to work
with multivariate learning machines, namely the learning machines which are
optimized during the training phase to handle multiple features simultaneously.
SVM for example, is a typical multivariate learning machine (Chen & Lin, 2006).
165
The F-score is a simple technique which estimates the discrimination of two sets of
real numbers. For a given number of training vectors O_, � = 1,2,… ,k if the
number of positive and negative instances are )2 and )/, respectively, then the F-
score of the &7� feature is defined by the following expression (Chen & Lin, 2006):
Ê�&� = �]̅|���/]̅|�²2�]̅|���/]̅|�²ùÕ��ù ∑ �]̅�,|���/]̅|����²2 ùÕ��ù∑ �]̅�,|���/]̅|����²Õ���ùÕ���ù (4.54)
where O̅\, O̅\�2�, O̅\�/� are the average of the &7� feature of the positive and negative
datasets, respectively. Similarly in equation (4.54), O̅_,\�2� is the &7� feature of the
�7� positive instance and O̅_,\�/� is the &7� feature of the �7� negative instance. The
numerator in equation (4.54) represents the discrimination between the positive
and negative sets whereas the denominator represents the one within each of the
two sets. In practice it is considered that, the larger the F-score is, the more likely
this feature is more discriminative. Since the F-score is a simple and effective
technique, the procedure for selecting optimum features (high F-scores) using the
SVM is summarized in the following steps (Chen and Lin, 2006):
1. Calculate the F-score of every feature.
2. Pick possible thresholds (using the naked eye) to cut low and high F-scores.
3. Then for each threshold, do the following:
a) Drop features with F-scores below this threshold.
b) Split the training data into two random sets: ºc�0\� and ºÌ0h\ü .
c) Set ºc�0\� be the new training data. Apply the technique discussed in
Section 4.2.4, to obtained a binary SVM classifier; use the classifier to
predict ºÌ0h\ü.
166
d) Repeat the steps above 10 times and then calculate the average
validation error for each trial; perform 10-fold CV.
4. Choose the best threshold, i.e., the threshold with the lowest average
validation error; lowest 10-fold CV accuracy.
5. Drop features with F-score below the selected threshold. Then apply the
Non-linear SVM in Section 4.2.4.
In the above procedure, possible thresholds can be determined by the human eye.
This step can be automated by gradually adding high F-score features, until the
validation accuracy of the model decreases.
4.4.2 SVM-RFE Using Random Forest
The F-score feature ranking technique discussed for SVM-RFE in Section 4.4.1 (in
equation (4.54)) is concerned with the removal of one feature at a time. Starting
with a large number of features, when the goal is to obtain a small subset of
relevant features, it is necessary to remove more than one feature at a time, which
can be accomplished using filtering. Filtering can be accomplished by applying the
Random Forest (RF) technique (Svetnik et al., 2004).
The Random Forest (RF) is a typical classification tool, but also provides feature
importance (Breiman, 2001). In the RF technique, a forest contains many decision
trees which constructed using instances with randomly sampled features. So, the
prediction using the RF is made by a majority vote of decision trees. In order to
obtain feature importance, the training set is split into two sets. Training using the
first set and predicting using the second set, an accuracy �+\� can be obtained. Also,
the values of the j7� feature are randomly permuted in the second set to obtained
167
another accuracy �+ª�. Thus, the difference between these accuracies �+\ and +ª� shows the importance of the j7� feature (Chen & Lin, 2006).
In practice the RF technique is known to have a high computational time. Thus,
before using the RF, a subset of features are selected at first using F-score feature
ranking technique indicated in Section 4.4.1. The SVM-RFE technique applied in
this research for the purpose of feature selection is referred to as “F-score + RF +
SVM”, which is summarized in the following steps (Chen & Lin, 2006):
1. F-score
a) Consider the subset of features obtained using the F-score feature
ranking technique in Section 4.4.1.
2. Random Forest (RF)
a) Initialize the RF dataset so as to include all training instances with
the subset of features selected from Step 1(a). Then use the RF
technique to obtain the rank of features.
b) Next, use the RF as a classification engine in order to perform 10-fold
CV on the working set.
c) Then update the working set by removing half of the less important
features and go to Step 2(b). Stop if the number of features is small.
d) Amongst the various feature subsets chosen above, select the feature
subset with the lowest 10-fold CV error.
3. Support Vector Machine (SVM)
a) Apply the technique discussed in Section 4.2.4, to obtained a binary
SVM classifier;
168
4.5 Summary
An overview of pattern recognition was presented in this chapter, with particular
emphasis on a specific machine learning technique, namely, the Support Vector
Machine (SVM). SVMs are used intensively in this research. The reason for using
SVM as the main machine learning technique for this research is discussed in
Sections 1.3 and 3.3.5. Section 4.1 presented some introductory notions regarding
the theoretical concepts of learning machines. Section 4.2 introduced the
fundamental concepts of the statistical learning theory and presented the
mathematical formulation of the SVM developed by Vapnik (1998) which describe
the statistical aspects of automated machine learning. Towards the end of this
chapter, Section 4.3 presented the theoretical concepts of ANNs whereas Section
4.4 discussed a Recursive Feature Elimination (RFE) technique used for the
selection of the optimal subset of texture features for the learning machine (SVM).
169
CHAPTER 5
FRAMEWORK MODELING
5.0 Overview
This chapter presents modeling of the framework (system) proposed in Chapter 1
for the classification of benign and malignant abnormalities in digital
mammograms. As discussed in Chapter 3, the proposed framework is composed of
two main techniques, namely, image processing and machine learning. The image
process and machine learning techniques identified in Section 3.3, are applied in
the proposed system in this chapter, which are discussed in Section 5.1 and 5.2.
The modeling of the system consists of three main stages, namely: Mammogram
Image Processing (Section 5.3), Texture Feature Extraction and Selection (Section
5.4) and Classification Engine (Section 5.5). Sections 5.3 through 5.5 describe each
stage in detail during the development of the proposed framework.
5.1 Proposed Framework
As discussed in Section 1.4, a prototype framework for the computerized detection
of breast cancer from digital mammograms is presented in Figure 1.4. The
modeling of this framework consists of three main stages, namely: Mammogram
Image Processing (see Section 5.3), Texture Feature Extraction and Selection (see
Section 5.4) and Classification Engine (see Section 5.5) as indicated in Figure 5.1.
In the first stage, image processing techniques and algorithms are applied on the
digital mammographic images for the purpose of image preprocessing (see Section
170
3.3.1) and image segmentation (see Section 3.3.2). In this research, mammogram
preprocessing includes: noise removal (using Median Filtering), background
suppression (using Global Thresholding) and artifact/wedge and label suppression
(using Morphological Operations) as shown in Figure 5.1.
During the mammogram segmentation stage, first the breast profile is optimally
segmented from the background (using Contrast Enhancement) so that the breast-
skin edges are retained in the breast profile (tissue). Secondly, the pectoral muscle
is suppressed (using Seeded Region Growing) from the mammograms, since it may
bias the procedures in the detection of malignant and benign tumors (Nicolaou et
al., 2008), (Xu et al., 2007), (Mirzaalian et al., 2007). The ROIs (abnormal regions)
in the segmented mammogram images are extracted using the Ground Truth (GT)
data from radiologists’ diagnosis of the mammographic datasets, as discussed in
Section 5.2.1.
In the second stage, texture features analysis is performed using Gray Level Co-
occurrence Matrices (GLCMs), which are employed in this research to compute the
texture features (see Section 3.3) from the ROIs (abnormal regions). Standard
GLCM texture descriptors proposed by Haralick et al. (1973) and Haralick (1979)
are used with other texture descriptors proposed by Soh and Tsatsoulis (1999)
and Clausi (2002). The optimal subset of texture features is selected using the
SVM-Recursive Feature Elimination (SVM-RFE) technique (using F-score + Random
Forest + SVM) technique, as discussed in Section 4.4.
171
Figure 5.1: Flowchart of the proposed computerized breast cancer detection
framework
BENIGN
MALIGNANT
Third module: Classification Engine
First module: Mammogram Preprocessing
Input: Digital
Mammograms
IMAGE SEGMENTATION
1. Breast Profile Segmentation: Contrast Enhancement
2. Pectoral Muscle Removal: Seeded Region Growing (SRG) 3. Region of Interest (ROI) Extraction: Ground Truth (GT) data
IMAGE PREPROCESSING
1. Noise Removal: 2D Median Filtering
2. Background Separation: Global Thresholding
3. Radiopaque Artifact Suppression: Morphological Operations
Second module: Texture Feature Analysis
TEXTURE FEATURE EXTRACTION AND SELECTION
1. Texture Feature Extraction: GLCMs
2. Texture Descriptors: (Haralick, 1973), (Soh & Tsatsoulis, 1999) and (Clausi, 2002)
3. Optimum Feature Selection: F-score, Random Forest
CLASSIFIER DEVELOPMENT AND OPTIMIZATION
1. Classifier: Non-linear SVM (Binary classification)
2. Classifier Training: Sequential Minimal Optimization (SMO)
3. Classifier Parameter Optimization: Grid Search method
4. Classifier Validation: Cross-Validation
5. Performance Evaluation: Confusion Matrix and ROC curve
Output: ROI of Segmented
Mammogram Images
Output: Optimum Subset of
Texture Features
Output: Classification of
Abnormality
172
The last stage in Figure 5.1 uses the optimal subset of texture features to construct
a classification engine (classifier) using the non-linear SVM approach presented in
Section 4.2.4. The SVM classifier used in this research performs binary
classification between malignant and benign ROIs. The SVM is implemented using
the Sequential Minimal Optimization (SMO) algorithm (see Section 4.2.5). The
memorization capability of the SVM classifier is estimated using the Cross-
Validation (CV) approach discussed in Section 4.1.3.2. The Grid-Search method
presented in Section 4.2.4 is used to optimize (fine-tune) the hyperplane
parameters for the non-linear SVM. The accuracy of the proposed classification
engine is evaluated using a confusion matrix (in Table 3.1) such as the sensitivity,
specificity, FPF and AUC computed for a medical diagnosis test, as discussed in
Section 3.3.6. The output of the proposed system classifies the tested samples
(ROIs) as malignant or benign.
5.2 Research Methodology and Implementation
The framework proposed in this thesis for the modeling of a computerized breast
cancer detection system (in Figure 5.1) is indicated in Figure 5.2. The first three
stages in Figure 5.2 will be discussed in this chapter, and the fourth stage which
presents the testing results of the modeled framework, are discussed in Chapter 6.
In this research, the image processing and texture analysis techniques are applied
using the MATLAB Image Processing and Statistics Toolboxes. The SVM is applied
in the proposed framework using the LIBSVM library (Chang & Lin, 2010).
173
Figure 5.2: Flowchart of the research framework
Start of
Research
Image Preprocessing
Data Acquisition
(Digital Mammograms)
Image Segmentation
SVM Training
SVM Parameter Optimization
Trained Model (Classification
Engine)
Classification
(SVM Testing)
Comparison with Machine Learning
techniques
Texture Feature Extraction
Optimal Feature Subset Selection
Mammogram Image
Processing
Texture Feature Extraction
and Selection
SVM Classification
Accuracy
Bad
Good
End of
Research
STAGE 1
Detection Results
Performance Evaluation
and
Comparison
STAGE 2
STAGE 3
STAGE 4
Classification Engine
Development
174
5.2.1. Data Acquisition
Digital mammograms are used as the standard inputs into the proposed
framework, as indicated in Figure 1.4 and Figure 5.1. The data used in this research
is obtained from two distinct sources:
1. Mammographic images obtained from the University of Malaya Medical
Centre (UMMC), Kuala Lumpur.
2. Mammography dataset obtained from the Mammographic Image Analysis
Society (MIAS) database (Suckling et al., 1994).
The mammography images obtained from the UMMC are used in this research as
the local dataset. In order to generalize the accuracy of the proposed system in
better terms, a second dataset, namely the MIAS database (Suckling et al., 1994), is
used in this research as the external dataset. Both datasets used in this research
consist of the mediolateral oblique (MLO) view of the mammograms.
Digital mammograms in the local dataset were acquired for Malaysian patients
treated at the Department of Radiology at the University Malaya Medical Centre
(UMMC) (Dept. of Biomedical Imaging at UMMC, 2010) in collaboration with the
Faculty of Medicine at the University of Malaya (Faculty of Medicine at UM, 2010).
The MIAS database of mammograms (Suckling et al., 1994) is a well-known
published image database of 322 digital mammograms from the Mammographic
Image Analysis Society, United Kingdom. The mammograms in the MIAS database
are taken from the United Kingdom National Breast Screening Programme and this
dataset has been cited in many peer reviewed research articles (Ferrari et al.,
175
(2004, 2001)), (Selvan et al., 2006), (Dua et al., 2009), (Özekes et al., 2005),
(Ibrahim et al., 1997), (Domínguez & Nandi, (2008, 2009a, 2009b)), (Sheshadri &
Kandaswamya, 2007), (Hassanien, 2007), (Subashini et al., 2010), (Song et al.,
2009). Table 5.1 indicates the number of mammogram samples acquired from the
UMMC and the MIAS dataset.
Table 5.1: Mammography data acquired from UMMC and MIAS database
Data Source Malignant
Samples
Benign
Samples
Normal
Samples
Database
Samples
University Malaya Medical Centre (UMMC)
58 46 156 260
mini-MIAS Database of Mammograms
(Suckling et al., 2004) 52 60 210 322
Total Samples 110 106 366 582
Both datasets in Table 5.1 consist of standard images of dense, fatty and fatty-
glandular breasts, which are classified into three major categories: malignant,
benign and normal. Samples of digital mammography images acquired from the
UMMC and the MIAS datasets are shown in Figures 5.3 and 5.4 respectively. Based
on the visual inspection of the mammography images, MCCs were found in most
malignant and benign cases apart from mass lesions.
The total number of mammographic images obtained from UMMC is limited, i.e., a
total of 256 mammographic images, as indicated in Table 5.1. The reason for this is
due to the fact that UMMC has only recently implemented digital mammography,
that is, in 2008. Thus, over a course of nearly two years from 2008 to 2010, only a
limited number of malignant and benign cases are available in digital format.
176
(a) (b)
(c) (d)
Figure: 5.3: Mammography images acquired from UMMC
(Dept. of Biomedical Imaging at UMMC, 2010)
The UMMC and MIAS mammography images are digitized at 200 micron pixel
edge, with a size of size of 1024 × 1024 pixels. Each pixel in the grayscale
mammogram image represents the pixel intensity in the range of [0, 255] (8-bit).
177
(a) (b)
(c) (d)
Figure: 5.4: Mammography images acquired from MIAS (Suckling et al., 2004)
Other information obtained for the mammogram images from both datasets
include the Ground Truth (GT) data and markings. The GT data contains the
diagnostic results of the radiologists’ interpretation of mammograms, i.e. the
location of benign and malignant abnormalities, as indicated in Figure 5.5. The GT
data for the UMMC dataset was provided by expert radiologists who have
diagnosed and treated Malaysian patients with benign and malignant
abnormalities. The GT data for the MIAS database (Suckling et al., 1994) was
178
acquired online together with the mammography dataset. The common
information retrieved from the GT data for both the UMMC and the MIAS dataset
are:
1. The centre of location of the abnormality (malignant/benign) in the digital
images in the format of �O, P� co-ordinates.
2. The approximate radius in pixels of a circle enclosing the abnormality area.
In this research, this GT data is identified for the purpose of extracting ROIs
(abnormal regions) from the mammography images. The following sections in this
chapter discuss the modeling of the proposed framework (system) as outlined in
Section 5.1.
(a) (b)
Figure: 5.5: Ground Truth (GT) markings by expert radiologists on acquired
mammography datasets
179
5.3 Mammogram Image Processing
5.3.1 Image Preprocessing
Digital mammograms typically contain artifacts in the form of labels, wedges and
markers in the background region. These artifacts are usually radiopaque such that
that they are not transparent to radiation. The major problem with the precise
segmentation of the breast region is due to the existence of such artifacts, which
may cause trivial segmentation algorithms to fail. The mammogram preprocessing
stage indicated in Figure 5.2 involves noise removal and radiopaque artifact
suppression in order to suppress the background (black pixels) in the
mammogram images. Another purpose of mammogram preprocessing is to
improve the reliability and robustness of the mammogram segmentation, as
discussed in the following sections.
5.3.1.1 Noise Removal
The grayscale mammography images are digitally represented using the MATLAB
Image Processing Toolbox (Image Processing Toolbox, 2010). The digital images
are mathematically represented using equation (3.6), where the range of intensity
values of the acquired mammogram images has [0,255] gray levels.
Digitization noises such as horizontal and vertical lines tend to appear on most of
the mammogram images, as shown by indication markers in Figure 5.6. These
noises are removed from the mammogram images by applying a Two-dimensional
Median Filtering approach (Lim, 1990) in a 3-by-3 connected neighborhood.
180
(a) (b)
(c) (d)
Figure: 5.6: Digitization noises (lines) in mammographic images
181
(a) (b)
(c) (d)
Figure: 5.7: Mammogram images after noise removal using 2D median filtering
182
In the 2D median filtering approach, each output pixel contains the median value
in the 3-by-3 neighborhood around the corresponding to the pixel of interest in the
image. The edges of the image however, are replaced by zeros (total absence or
black color). This does not affect the original image, since the ROI of the image does
not include the edges or boundaries of the image. Figure 5.7 shows the
mammogram images in Figure 5.6 after applying the 2D median filtering approach.
The digitization noises, i.e. the horizontal and vertical lines are removed from the
mammogram images without affecting the breast profile.
5.3.1.2 Radiopaque Artifact Suppression
Mammogram artifacts such as identification labels, markers, and wedges are
radiopaque such that that they are not transparent to radiation. These artifacts are
small emulsion continuity faults on the mammogram films which look like
calcifications. Suppressing radiopaque artifacts and the background region within
a mammogram image, increases the region homogeneity and also improves the
reliability and robustness of the breast profile separation. A mammogram artifact
suppression algorithm based on the area morphology presented by Wirth et al.
(2004) is adopted in this research (Yapa & Harada, 2008), (Wirth et al., 2007).
In order to use the area morphology approach presented by Wirth et al. (2004), the
grayscale [0,255] mammogram image needs to be transformed into the binary
[0,1] format. The simplest technique for transforming a grayscale image into
binary is by using thresholding (see Section 3.4.3.1). In order to convert a
grayscale image into binary, a grayscale threshold for that image needs to be
determined in order segment the artifacts and the background, while keeping the
breast-skin edges in contact, so as not to lose information from the breast profile.
183
In background and artifact suppression, a global threshold ��� is a user-determined
value that is used to optimally segment the background region and the radiopaque
artifacts from the breast profile for a mammogram image dataset. In order to
determine a global threshold ��� for a mammogram image dataset, a trial and error
procedure is typically used, where the segmentation performance of all the
mammogram images is evaluated for all possible threshold levels. In this research,
visual inspection of the segmented mammogram images determines the global
threshold to be: � = 18, for all possible threshold levels in between 0 to 255.
Figure 5.8(a) and Figure 5.8(c) represent the histograms of the mammogram
images in Figure 5.9(a) and Figure 5.9(c) respectively. Figure 5.9(b) and Figure
5.9(d) show the resulting histograms obtained after thresholding the images
corresponding to Figures 5.9(a) and 5.9(d) respectively. As observed from Figures
5.8 and 5.9, applying a threshold value of � = 18 sets all the pixels within the
intensity range of 0 to 18 to a single value of 0 (background pixel).
The resulting binary images after global thresholding are indicated in Figure 5.9,
where the breast profile is segmented from the background region. Next,
radiopaque artifacts are suppressed by applying Morphological Operations (in
Section 3.4.4) on the binary images as illustrated in Figure 5.11. The artifact
suppression algorithm by Wirth et al. (2004) is modified in order to suit both the
local and external datasets and is presented as follows:
Mammogram Artifact Suppression Algorithm
1. All objects present in the binary image are labeled. The binary objects
consist of the radiopaque artifacts and the breast profile as indicated in
Figure 5.11(b).
184
2. The ‘area’ (actual number of pixels in a region) is calculated for all objects
(regions) in the binary image as shown in Figure 5.11(b).
(a) (b)
(c) (d)
Figure: 5.8: Histograms after applying global thresholding using � = 18.
(a) Original histogram of mammogram image in Figure 5.7(a)
(b) Histogram of mammogram image in Figure 5.7(a) after thresholding.
(c) Original histogram of mammogram image in Figure 5.7(d)
(d) Histogram of mammogram image in Figure 5.7(d) after thresholding.
3. The largest object (area calculated in Step(2)) in the binary image in Figure
5.11(b) is then selected. This operation opens a binary image and removes
all objects in the binary image, retaining the largest object, which is the
breast profile region as shown in Figure 5.11(c). This operation uses an 8-
connected neighborhood for a two-dimensional connective.
185
(a) (b)
(c) (d)
Figure: 5.9: Thresholding for segmentation of the breast profile region from the
background region. (a) Original mammogram image
(b) Mammogram image in Figure 5.9(a) after breast profile separation
(c) Original mammogram image
(d) Mammogram image in Figure 5.9(c) after breast profile separation
4. Next, an operation to remove isolated pixels (individual 1’s that are
surrounded by 0’s) and reduce distortion is applied the binary image. This
algorithm checks all pixels in the binary image and if the 8-8 connected
186
neighborhood around the pixel of interest (pixel with value 1) consists of all
0’s (isolated pixel), then the pixel of interest is set as 0 (background). An
example of an isolated pixel is given by matrix º:
º = �0 0 00 1 00 0 0� (5.1)
5. Another operation is applied to the binary image to smoothen less visible
noise. This simple algorithm checks all pixels in the binary image and sets a
pixel to 1 if five or more pixels in its 3-by-3 neighborhood are 1's,
otherwise, it sets the pixel to 0.
Figure: 5.10: Flat disk-shaped morphological structuring element (STREL)
6. The binary image is eroded (see Section 3.4.4.2) using a flat, disk-shaped
morphological structuring element (STREL) as shown in Figure 5.10. Using
a trial and error procedure, the optimum value of the radius �� for the
acquired mammographic image dataset is found to be = 3.
7. Next, the binary image is dilated (Section 3.4.4.1), using the same STREL
object adopted in Step (6).
187
8. The holes in the binary image are filled using a simple algorithm to fill all
holes in the binary image, where a hole is defined as a set of background
pixels that cannot be reached by filling in the background from the edge of
the image.
(a) (b)
(c) (d)
Figure 5.11: Suppression of radiopaque artifacts
(a) Original grayscale image with artifact and label
(b) Thresholded image using a value of T = 18 ������ = 0.0706� (c) Selection of the largest object with respect to Area
(d) Grayscale image with radiopaque artifacts suppressed
188
9. The binary image obtained in Step (8), (Figure 5.11(c)), is multiplied with
the original mammogram image in Figure 5.11(a) to obtain Figure 5.11(d).
To illustrate this process, consider the binary image obtained from Step (8)
(see Figure 5.11(c)) as O and the original mammogram image as P (see
Figure 5.11(a)). Then each element in array O is multiplied with the
corresponding element in array P to return the result in the corresponding
element of the output array ~, which is the grayscale image with the
background and the radiopaque artifacts suppressed, as shown in Figures
5.11(d) and 5.12(b).
(a) (b)
Figure 5.12: Segmented breast profile region
(a) Binary image of thresholded mammogram
(b) Grayscale image after background and artifact suppression
The grayscale image obtained in Step (9) as indicated in Figure 5.12(b), has
radiopaque artifacts suppressed and the breast profile region separated from the
189
background. The grayscale images resulting from this process are further used to
perform mammogram segmentation as indicated in the following section.
5.3.2 Image Segmentation
The presence of the pectoral muscle in mammogram images effects the results of
intensity based image processing methods and can bias procedures in the
detection of masses and MCCs, as discussed in Section 3.3.2. Several research
groups indicated that during the computerized detection of masses and MCCs the
pectoral muscle should be excluded from processing (Kwok et al. 2004), (Raba et
al. 2005), (Nicolaou et al. 2008), (Xu et al. 2007), (Ferrari et al. 2004), (Mirzaalian
et al. 2007), (Bajger et al. 2005). In order to confirm these findings, the GT data of
the mammograms acquired in this research was analyzed, which indicated that no
mass lesions or MCCs appear in the pectoral muscle region. This justifies the fact to
remove the pectoral muscle from the mammogram images.
The Seeded Region Growing (SRG) technique (in Section 3.4.3.3.2) as described in
the framework in Figure 5.1, is proposed in this research as a suitable technique
for the purpose of pectoral muscle segmentation. In this research, the pectoral
muscle segmentation is implemented in MATLAB, which consists of the following
four stages:
Stage 1 ― Determining Orientation of the Breast Pro@ile
The breast orientation (left-side or right-side) in each mammogram image needs to
be determined prior to performing Seeded Region Growing (SRG) for pectoral
muscle segmentation. In order to perform SRG, a seed needs to be placed inside the
pectoral muscle, hence determining the breast orientation is crucial.
190
(a) (b) (c)
(d) (e) (f)
Figure 5.13: Cropping breast profile in mammogram images to image borders
(a) Binary image with right oriented breast
(b) Binary image in Figure 5.13(a) cropped from the left and right
(c) Binary image in Figure 5.13(a) cropped from the top and bottom
(d) Binary image with left oriented breast
(b) Binary image in Figure 5.13(d) cropped from the left and right
(e) Binary image in Figure 5.13(d) cropped from the top and bottom
When the pectoral muscle is located on the top left corner (see Figure 5.7(a)) of the
mammogram the breast points towards the right side, hence the breast is termed
as right oriented. Similarly, when the pectoral muscle is located on the top right
191
corner (Figure 5.7(b)) the breast points towards the left side, hence the breast is
left oriented. A simple algorithm for determining the breast profile orientation is
developed in this research, which is as follows:
(a) In order to determine the breast profile orientation (left-side or right-side)
of mammograms using an automated procedure, the binary image in Figure
5.12(a) is used. As indicated in Figure 5.13(b) the binary image is cropped
left to right and top to bottom, such that the breast profile object touches all
four borders (left, right, top and bottom) of the binary image, as shown in
Figure 5.13(c).
(b) After the breast profile is cropped, the sum of the pixel intensities in first
and last five columns of the binary images in Figure 5.13(c) and Figure
5.13(f) is calculated. To determine the breast profile orientation, intensities
of the pixels in the shaded area of equations (5.2) and (5.3) are calculated,
which represent the pixels in first and last five columns of the binary image
respectively.
� =QRRS
��0,1� ��0,2� ��0,3� ��0,4� ��0,5� … ��0,U���1,1� ��1,2� ��1,3� ��1,4� ��1,5� … ��1,U�⋮ ⋮ ⋮ ⋮ ⋮ ⋮��W, 1� ��W, 2� ��W, 3� ��W, 4� ���,�� … ��W,U�XYYZ (5.2)
� =QRRRS ��0,1� … ��0,U − 4� ��0,U − 3� ��0,U − 2� ��0, U − 1� ��0, U���1,1� … ��1,U − 4� ��1,U − 3� ��1,U − 2� ��1, U − 1� ��1, U�⋮ ⋮ ⋮ ⋮ ⋮ ⋮��W, 1� … ���,� − �� ��W,U − 3� ��W, U − 2� ��W, U − 1� ��W,U�
XYYYZ (5.3)
192
In equations (5.2) and (5.3), the binary images have W rows and U
columns, or in other words, the images are of W ×U dimensions. Each
element of the matrix ��W,U� in equations (5.2) and (5.3) corresponds
to an image pixel, represented by a binary value [0,1]. Using the shaded
area in equation (5.2) and equation (5.3), the sum of the first and last 5
columns of the binary images in Figures 5.13(c) and 5.13(f) is calculated
respectively, using the following expressions:
%ík�\��c =�������
sum���0,1�, ��1,1�, … , ��W, 1��+sum���0,2�, ��1,2�, … , ��W, 2��+sum���0,3�, ��1,3�, … , ��W, 3��+sum���0,4�, ��1,4�, … , ��W, 4��+sum���0,5�, ��1,5�, … , ��W, 5�� ��!��"
(5.4)
%íkh0�c =�������
sum���0, U − 4�, ��1, U − 4�, … , ��W, U − 4��+sum���0, U − 3�, ��1, U − 3�, … , ��W, U − 3��+%um���0, U − 2�, ��1, U − 2�, … , ��W,U − 2��+sum���0, U − 1�, ��1, U − 1�, … , ��W, U − 1��+sum���0, U�, ��1, U�, … , ��W,U�� ��!��"
(5.5)
where %ík�\��c and %íkh0�c are the sums (as integer values) of the
binary values [0,1] in the first and last five columns of the binary image
in Figure 5.13(c) and Figure 5.13(f).
(c) Next, the sums of the binary values (%ík�\��c and %íkh0�c) are used to
construct a simple IF-THEN rule for identifying the breast profile as left-
sided or right-sided. The decision making rule developed in this research is
as given below.
193
Pseudo-code of for determining the breast orientation
if �%ík�\��c > %íkh0�c�
%Breast is oriented towards the right
%See Figure 5.13(c)
elseif �%ík�\��c < %íkh0�c� %Breast is oriented towards the left
%See Figure 5.13(f)
end if
The IF-THEN logic in the pseudo-code above compares the sum of the intensity
values in the binary image, namely %ík�\��c and %íkh0�c and accurately identifies
the orientation of the breast profile of all mammogram images acquired (in Table
5.1). Using this rule, Figure 5.13(c) is identified as a right-orientated breast, while
Figure 5.13(f) is identified as a left-orientated breast.
Stage 2 ― Breast Pro@ile Contrast Enhancement
Prior to pectoral muscle segmentation, the contrast of the breast profile,
particularly the pectoral muscle region, needs to be enhanced. Since the pectoral
muscle contains the majority of the brightest pixels in the breast profile, hence the
contrast of the breast profile needs to be enhanced such that the grayscale pixels in
the pectoral muscle become brighter. By using contrast enhanced images, the SRG
algorithm can more reliably identify the pectoral muscle border from the breast
profile.
In order to perform contrast enhancement, the mammogram image obtained in
Figure 5.12(b) is used. Contrast enhancement is performed by finding the limits to
contrast stretch an image, where the tolerance level is a scalar quantity, which
194
saturates fractions at low and high pixel levels. In this research, the default
tolerance value of ' = k&) k+O� is used, where k&) is the smallest grayscale
value in the mammogram image and k+O is the highest grayscale value in the
mammogram image.
(a) (b)
(c) (d)
Figure 5.14: Contrast enhancement of a mammogram image
(a) Mammogram image obtained after mammogram preprocessing (Figure 5.12).
(b) Histogram of the original image in (a)
(c) Contrast enhancement applied to the mammogram image in (a)
(d) Histogram of contrast enhanced image in (c)
195
Figure 5.14 shows the contrast enhancement technique applied to the
mammogram images. As indicated in Figure 5.14(d) the histogram of the original
image in Figure 5.14(b) is stretched, which increases the number of brighter pixels.
Stage 3 ― Seeded Region Growing
After the breast orientation is determined in the first stage and the breast profile
contrast is enhanced in the second stage, the Seeded Region Growing (SRG)
technique (Adams & Bischof, 1994) is then applied to segment the pectoral muscle
as discussed in Section 3.4.3.3.2. In order to implement SRG a seed needs to be
placed inside the pectoral muscle of the grayscale mammogram image.
Using results obtained from the first and second stages, if the breast profile is
right-orientated a seed is placed in the top-left corner (pectoral muscle) of the
image shown in Figure 5.15(a), corresponding to pixel ��W, 5� circled in equation
(5.2). Similarly, if the breast profile is right-orientated, a seed is placed in the top-
right corner (pectoral muscle) of the mammogram image shown in Figure 5.15(d),
corresponding to pixel ��W, U − 4� circled in equation (5.3)). The theoretical
concepts of the SRG algorithm have been presented in detail in Section 3.4.3.3.2 of
this thesis. The following four steps are applied in the SRG process:
After the breast orientation is determined in the first stage and the breast profile
contrast is enhanced in the second stage, the Seeded Region Growing (SRG)
technique (Adams & Bischof, 1994) is next applied as discussed in Section 3.4.3.3.2
for segmenting the pectoral muscle. In order to implement SRG a seed needs to be
placed inside the pectoral muscle of the grayscale mammogram image. Using
results obtained from the first and second stages, if the breast profile is right-
196
orientated a seed is placed in the top-right corner (pectoral muscle) of the image
shown in Figure 5.15(a), else if the breast profile is right-orientated, a seed is
placed in the top-left corner (pectoral muscle) of the image shown in Figure
5.15(d). Theoretical concepts of the SRG algorithm are presented in detail in
Section 3.4.3.3.2 of this thesis. The following four steps are applied in the SRG
process:
a. The region is iteratively grown by comparing all unallocated neighboring
pixels to the region.
b. The difference between the pixel of interests’ intensity value and the
region’s mean is used as a measure of similarity.
c. The pixel with the smallest difference measure is allocated to the respective
region.
d. The process stops when the intensity difference between the region mean
and the new pixel become larger than the threshold value (maximum
intensity distance).
Using a trial and error procedure for inspecting the segmentation performance of
all the mammogram images for all possible threshold levels in between 0 to 255,
the optimum SRG threshold satisfying all mammogram images (in Table 5.1) is
determined to be ³ = 18. The SRG threshold is generally referred to as the
maximum intensity distance satisfying all mammogram images to reliably segment
the pectoral muscle from the breast profile region. The results of the segmented
pectoral muscle obtained after the SRG process, is a binary image as shown in
Figure 5.15(c) and Figure 5.15(f).
197
(a) (b) (c)
(d) (e) (f)
Figure 5.15: Segmentation of pectoral muscle using Seeded Region Growing
(a) Contrast enhanced breast profile right-orientated
(b) Binary image of Figure 5.15(a) showing separated breast profile
(c) Segmented pectoral muscle of Figure 5.15(a) using SRG
(d) Contrast enhanced breast profile left-orientated
(e) Binary image of Figure 5.15(d) showing separated breast profile
(f) Segmented pectoral muscle of Figure 5.15(d) using SRG
198
Next, the binary object in Figure 5.15(c) and Figure 5.15(f) is subtracted from the
breast profile region in Figure 5.15(b) and Figure 5.15(e) respectively. To illustrate
this concept, consider the breast profile in the binary image as O and the
segmented pectoral muscle in the binary image as P. Then each element in array P
is subtracted from the corresponding element in array O returning the difference in
the corresponding element of the output array ~, which is a binary image of the
breast profile with the pectoral muscle segmented, as indicated in Figures 5.16(a)
and 5.16(b). This process can be implemented in MATLAB by using the available
built in library such as the &k%í�'b+î' function.
Typically there is only one binary object present in the binary image (see Figure
5.16(a) and Figure 5.16(b)) after the segmentation of the pectoral muscle.
However in a few cases, other smaller objects may be present, i.e. parts of the
pectoral muscle near the breast profile and pectoral muscle border are retained. In
order to cater for these small binary objects, which represent the segmentation
noise, the following three steps are applied:
a. All objects in the binary image in Figures 5.16(a) and 5.16(b) are labeled.
b. The ‘area’ (actual number of pixels in a region) is calculated for all objects
(regions) in the binary images (in Figure 5.16(a) and Figure 5.16(b)).
c. The largest object (area calculated in Step(b)) in the binary images in Figure
5.16(a) and Figure 5.16(b) is selected. This operation opens a binary image
and removes all objects in the binary image, retaining the largest object,
which is the pectoral muscle as shown in Figures 5.15(c) and 5.15(f). This
operation uses an 8-connected neighborhood.
199
(a) (b)
Figure 5.16: Suppression of pectoral muscle from breast profile region
(a) Subtraction of binary image Figure 5.15(c) from binary image Figure 5.15(b)
(b) Subtraction of binary image Figure 5.15(e) from binary image Figure 5.15(f)
(a) (b)
Figure 5.17: Binary images of the segmented breast profile
(a) Right-oriented breast profile after morphological operations
(b) Left-oriented breast profile after morphological operations
200
Next, the morphological operations from Step (4) through Step (8) of Section
5.3.1.2 are performed on the binary images in Figure 5.16 to reduce distortion and
noise, fill holes and remove small objects. The resulting binary image after
performing the morphological operations is shown in Figure 5.17. Comparing
Figure 5.16 with Figure 5.17, it is observed that after applying morphological
operations to the binary image, the boundary between the breast profile region
(object) and the segmented pectoral muscle becomes smoother.
Stage 4 ― Smoothening Segmented Pectoral Muscle Boundary
Since the boundary between the breast profile region and the segmented pectoral
muscle contains rough (unsmooth) edges as indicated in Figure 5.17, thus a simple
and effective technique is applied to smooth the boundary of the segmented
pectoral muscle border. The pectoral muscle can be viewed as a right-angle
triangle in Figure 5.15(c) and Figure 5.15(f), which is illustrated in Figure 5.18(a)
and Figure 5.18(b) respectively.
In order to smooth the pectoral muscle boundary, i.e., the hypotenuse in Figure
5.18, the two points �O�, P�� and �O�, P�� in Figure 5.18 need to be determined in
Figure 5.17, with respect to the origin �0,0�. The two points �O�, P�� and �O�, P�� in
the format �b$Ù, î$�� in the digital images are determined by iterating through the
columns and rows of the binary image expressed in equation (3.6). The algorithm
proposed for the detection of the two points �O�, P�� and �O�, P�� for right and left
sided breasts is as follows.
201
Right-Oriented Breast
(a) To determine the point �O�, P�� in Figure 5.17(a), the first row of the binary
image ���0,1�, ��0,2�,… , ��0,U��inequation�5.2� of Figure 5.17(a) is
checked for the occurrence of a 1, since 0 represents black (background)
pixels. The first occurrence of 1 in the row is selected as the point �O�, P�� in
Figure 5.18(a).
(a) (b)
Figure 5.18: Pectoral muscle viewed as a right angle triangle
(a) Pectoral muscle in Figure 5.15(c) viewed as a right angle triangle for a
right-oriented breast profile
(b) Pectoral muscle in Figure 5.15(f) viewed as a right angle triangle for a
left-oriented breast profile
(b) Similarly, in order to find the point �O�, P�� in Figure 5.17(a), the first
column of the binary image ���0,1�, ��1,1�,… , ��W, 1�� in equation (5.2) of
Hypotenuse (c)
Opposite (a) Opposite (a)
Adjacent (b)
Adjacent (b)
90° 90°
c
a a
b b c
Hypotenuse
(c)
202
Figure 5.17(a) is checked for the occurrence of a 1. The first occurrence of 1
in the column is selected as the point �O�, P�� in Figure 5.18(a).
Left-Oriented Breast
a. To find the point �O�, P�� in Figure 5.17(b), the first row of the binary image
���0,U�, ��0,U − 1�,… , ��0,2�, ��0,1�� in equation (5.3) of Figure 5.17(b) is
checked in reverse order for the occurrence of a 1, since 0 represents black
(background) pixels. The first occurrence of 1 in the row is selected as the
point �O�, P�� in Figure 5.18(b).
b. Similarly, in order to find the point �O�, P�� in Figure 5.17(b), the last
column of the binary image ���0,U�, ��1,U�,… , ��W, U�� in equation (5.3)
of Figure 5.17(b) is checked for the occurrence of a 1. The first occurrence
of 1 in the column is selected as the point �O�, P�� in Figure 5.18(b).
After determining the two points �O�, P�� and �O�, P�� for the right angle triangle in
Figure 5.18, an offset value b is specified in the two points, where b is an integer
value in the range of 0 < b < 30. For right-sided breasts, specifying the offset value
b, transforms the two points �O�, P�� and �O�, P�� into �O� − b, P�� and �O�, P� + b�.
Similarly for left-sided breasts, specifying the offset value b, transforms the two
points �O�, P�� and �O�, P�� into �O� + b, P�� and �O�, P� + b�. Next, the straight line
equation is used to represent the hypotenuse of the right angle triangle in Figure
5.18, which is the boundary between the breast profile region and the segmented
pectoral muscle. A straight line is represented by the following expression:
P = kO + î (5.6)
203
where k is the slope (gradient) determining the steepness of the line, and î is the
intercept where the line crosses the y-axis. Using equation (5.6), the standard
equation of a straight line, the value of the slope k is calculated using the following
expression:
k = �P� − O��/�P� − O�� (5.7)
where the î-intercept is calculated using the following expression:
î = P� − �k ∗ P�� (5.8)
The equation of the straight line (in equation 5.6) can be found using the MATLAB
implementation. The following steps indicate how the equation of the straight line
is determined.
1. The equation of the straight line is found by specifying a vector for O in
equation (5.6), such that the vector O consists of P� points in the range of
O� < O < P�, where O� is always taken as 1. Using the vector O (number of
points in P�) with the determined values of k and î from equations (5.7)
and (5.8), equation (5.6) which is the standard equation of a straight line is
evaluated to form a straight line vector P, in order to represent the
hypotenuse of the right angle triangle in Figure 5.18.
Since the straight line obtained in vector P consists of floating point values,
hence the values are rounded off to the nearest integer. The process of
rounding off can be implemented in MATLAB by using the available built in
library such as the b$í)� function.
204
2. As the straight line (hypotenuse in Figure 5.18) needs to be superimposed
on the boundary between the breast profile region and the segmented
pectoral muscle in Figure 5.17, thus, vectors O and P (in equation (5.6))
calculated in Step 1 are combined together to form a data matrix of the
straight line. The process of matrix concatenation can be implemented in
MATLAB by using the available built in matrix operations.
The data matrix of the straight line obtained is superimposed on Figure 5.17
in order to smooth the boundary between the breast profile region and the
segmented pectoral muscle. The technique for superimposing the straight
line on Figure 5.17 is different for left and right oriented breasts.
For right-oriented breasts, the data matrix obtained in Step 2 is used with
the binary image in Figure 5.17(a) in order to smooth the boundary
between the breast profile region and the segmented pectoral muscle. To
illustrate this process in simple terms, after the straight line is
superimposed on the binary image in Figure 5.17(a) using the data matrix,
all pixel values on the left side of the straight line are set to 0 (background
pixel), which results in a smooth pectoral muscle boundary as indicated in
Figure 5.19(a).
Similarly, for left-oriented breasts, the data matrix obtained in Step 2 is
used with the binary image in Figure 5.17(b) in order to smooth the
boundary between the breast profile region and the segmented pectoral
muscle. To illustrate this process, after the straight line is superimposed on
the binary image in Figure 5.17(b) using the data matrix, all pixel values on
205
the right side of the straight line are set to 0 (background pixel), which
results in a smooth pectoral muscle boundary as indicated in Figure
5.19(b).
(a) (b)
Figure 5.19: Binary images after pectoral muscle boundary straightening
(a) Right-oriented breast profile applying straight line
(b) Left-oriented breast profile applying straight line
The erosion and dilation operations from Step (6) and Step (7) of Section 5.3.1.2
are performed on the binary images in Figure 5.19 to remove small objects formed
after superimposing the straight line. Finally, the binary image obtained in from
Step 1 is multiplied with the original mammogram image in Figure 5.12(b), the
result of which produces Figure 5.20.
206
The segmented mammograms obtained after suppressing the pectoral muscle
using this technique, are indicated in Figure 5.21(a) and Figure 5.21(c), where the
histograms are shown in Figure 5.21(b) and Figure 5.21(d) respectively. From the
histograms in Figure 5.21(b) and Figure 5.21(d), it is observed that removing the
pectoral muscle in the breast profile reduces the amount of brighter pixels in the
grayscale image. This is because the majority of brighter pixels in the breast profile
contribute to the pectoral muscle.
(a) (b)
Figure 5.20: Pectoral muscle segmentation from mammogram
(a) Pectoral muscle segmentation of right-oriented breast profile
(b) Pectoral muscle segmentation of left-oriented breast profile
207
(a) (b)
(c) (d)
Figure 5.21: Grayscale image after pectoral muscle segmentation
(a) Grayscale image of right-orientated breast profile
(b) Image histogram of Figure 5.21(a)
(c) Grayscale image of left-orientated breast profile
(d) Image histogram of Figure 5.21(c)
208
5.4 Texture Feature Extraction and Selection
5.4.1 Region of Interest (ROI) Selection
Features (heuristics) cannot be directly computed from the segmented
mammogram images obtained in Figure 5.21, since they will bias the detection
results. So, features need to be computed only from the abnormal (malignant and
benign) regions of the breast profile, while excluding all other unimportant parts
of the breast tissue. In this case, the GT data (in Section 5.2.1) acquired with the
mammography datasets is used in order to extract the Region of Interests (ROIs)
or abnormal regions (malignant and benign cells) from the segmented breast
profile in Figure 5.21(a) and Figure 5.21(c).
The information obtained for the malignant and benign cases from GT data and
markings as presented in Section 5.2.1 is:
1. The centre of location of the abnormality (malignant/benign) in the images
in �O, P� co-ordinates.
2. The approximate radius in pixels of a circle enclosing the abnormality area.
In this research, the location of the centre of the abnormality is used for the
purpose of ROI extraction, where the ROI is defined as an abnormality (malignant
or benign) in the mammogram. Since GLCM texture features can only be computed
for image data represented in the form of a 2D matrix having a fixed shape of U
rows and W columns, for this purpose, the ROI is selected in the shape of a square,
rather than a circle as in the GT data.
209
Figure 5.22: Extraction of samples using different “square” ROI sizes
The ROIs of the segmented mammogram images in Figure 5.21 are extracted using
Figure 5.22 such that the centre of location of the abnormality i.e., the �O, P� co-
ordinates in Figure 5.22 are the origin pixel �0,0� of the ROI in the mammogram. In
this research, to estimate the size of mammographic abnormalities using a square
ROI (see Figure 5.22), it is more relevant to consider the diameter of the
abnormality. Since, the diameter is two times the radius, thus, different sizes of the
diameter are used in this research, in order to determine the most optimum ROI
size.
Through the analysis of the GT data, the minimum and maximum diameter in
pixels of a circle enclosing all malignant and benign abnormalities is found to be 48
and 130 pixels respectively. Using this information the most common ROI sizes
enclosing the majority of malignant and benign abnormalities are determined from
the GT data, which are: 48 × 48 pixels, 64 × 64 pixels, 96 × 96 pixels, 110 × 110
pixels, 128 × 128 pixels, 136 × 136 pixels and 148 × 148 pixels.
rtop
rleft rright
rbottom
210
The �O, P� co-ordinates of the centre of the abnormality represent the ROI origin
�0,0� or the ROI centre. In order to extract ROIs, the �O, P� co-ordinates from the GT
data need to be represented by using integers as offset values.
As observed from Figure 5.22, the offset values representing the size of the square
ROI are: bc�� , bà�cc�� , bh�c and b�\�dc. The offsets used to extract ROIs of size 48 ×
48 pixels are: bc�� = �O, P + 23�, bà�cc�� = �O, P − 24�, bh�c = �O − 24, P� and
b�\�dc = �O + 24, P�. Similarly, to extract ROIs of size 64 × 64 pixels, the offsets
used are: bc�� = �O, P + 31�, bà�cc�� = �O, P − 32�, bh�c = �O − 32, P� and
b�\�dc = �O + 31, P�. The same procedure applies for the offsets of other ROI sizes.
(a) (b) (c)
(d) (e) (f)
Figure 5.23: ROIs of benign abnormalities (from labeled GT data)
(a) Calcification (b) Circumscribed mass
(c) Spiculated mass (d) Ill-defined mass
(e) Architectural distortion (f) Asymmetrical mass
211
Malignant and benign ROIs extracted from the acquired mammography images
using the technique in Figure 5.22 are shown in Figure 5.23 and Figure 5.24. As
observed from Figures 5.23 and 5.24, the ROIs with bright spots (or small clusters)
indicate MCCs and the shape of the masses (lesions) in the ROIs is categorized
using the BI-RADS lexicon presented in Section 2.6.4.1. The total number of ROIs
extracted using the GT data with the acquired mammography datasets is shown in
Table 5.2.
(a) (b) (c)
(d) (e) (f)
Figure 5.24: ROIs of malignant abnormalities (from labeled GT data)
(a) Calcification (b) Circumscribed mass
(c) Spiculated mass (d) Ill-defined mass
(e) Architectural distortion (f) Asymmetrical mass
5.4.1.1 Necessity for Mammogram Image Processing
For the purpose of the ROI extraction (see Figure 5.22), the necessity for
mammogram image processing and segmentation as discussed in Sections 5.3.1
and 5.3.2, arises. One of the main concerns during the ROI extraction is that a
212
minority of ROIs include the background region (black pixels), which can bias the
texture feature extraction process. This happens in mammograms normally when
benign and malignant masses/MCCs are present near the edges of the segmented
breast profile as indicated in Figure 5.25.
Table 5.2: ROIs extracted from acquired mammography datasets in Table 5.1.
Data Source Malignant
ROIs
Benign
ROIs
Database
ROIs
University Malaya Medical Centre (UMMC)
64 48 112
mini-MIAS Database of Mammograms
(Suckling et al., 2004) 54 66 120
Total ROIs (Samples) 118 114 232
In order to extract optimum texture features using GLCMs from the ROIs in Figure
5.25, background pixels in the ROIs need to be separated and the pectoral muscle
needs to be segmented from the breast profile region. The theoretical foundation
and the practical implementation using algorithms and techniques for background
separation and pectoral muscle segmentation have been thoroughly described in
Section 3.2 and presented in Sections 5.3.1 and 5.3.2.
In this research, for the purpose of texture feature extraction, only the segmented
breast profile region in the ROIs, indicated in Figure 5.26 is used. The ROIs in
Figure 5.26 show that the background region (pixels with an intensity value of 0)
in the breast profile is excluded from the texture feature extraction process. The
reason why the background region is excluded from texture feature extraction is
because GLCMs are restricted to perform calculation for counting occurrences of
image pixels having an intensity value of 0.
213
(a) (b) (c)
(d) (e)
Figure 5.25: ROIs containing unsegmented background region (from labeled GT
data)
(a) Benign ROI (b) Benign ROI (c) Malignant ROI
(d) Malignant ROI (e) Malignant ROI
5.4.2 Texture Feature Extraction
For the purpose of texture feature extraction the malignant and benign ROIs
obtained in Figure 5.23 and Figure 5.24 are used. GLCMs are known to be the most
common and successful techniques for texture analysis of digital mammograms as
illustrated in Table 3.3 and Section 3.5.3.
The theoretical background of GLCMs is presented in Section 3.5.3.1.1 of this
thesis, whereas Section 3.5.3.1.2 presents the standard GLCM texture descriptors
proposed by Haralick et al. (1979) as shown in Table 3.4. In this research apart
from using standard GLCM texture descriptors discussed by (Haralick, 1973),
other recent GLCM texture descriptors discussed by Clausi (2002), Soh &
214
Tsatsoulis (1999) and the MATLAB Image Processing Toolbox are adopted for
texture feature computation, as shown in Tables 5.3 through 5.5 respectively.
(a) (b) (c)
(d) (e)
Figure 5.26: ROIs in Figure 5.25 with segmented background region
(a) Benign ROI (b) Benign ROI
(c) Malignant ROI (d) Malignant ROI (e) Malignant ROI
Table 5.3: GLCM texture descriptors from Clausi (2002)
No. Texture Descriptor Formula Equation
No.
1. Inverse Difference Normalized
∑ û�\,ª��2|\/ª| where î�&, j� is the co-occurrence probability between grey levels �&andj� defined as: î�&, j� =
��\,ª�∑ ��\,ª�%|,&�ù
(5.9)
2. Inverse Difference Moment Normalized
∑ '|&�2�\/ª�² (5.10)
215
Table 5.4: GLCM texture descriptors from Soh & Tsatsoulis (1999)
No. Texture Descriptor Formula Equation
No.
1. Autocorrelation
∑ ∑ �&j�#�&, j�ª\ where #�&, j� represents the number of occurrences of grey levels �&andj�. (5.11)
2. Cluster Prominence
∑ ∑ �& + j − °] − °z��ª\ #�&, j�
where #�&, j� is the �&, j�th entry in a normalized GLCM. The mean for the rows and columns of the matrix are: °] = ∑ ∑ & ∙ª\ #�&, j� °z = ∑ ∑ j ∙ª\ #�&, j�
(5.12)
3. Cluster Shade ∑ ∑ �& + j − °] − °z��ª\ #�&, j� (5.13)
4. Dissimilarity ∑ ∑ |& − j| ∙ #�&, j�ª\ (5.14)
5. Homogeneity (Soh & Tsatsoulis)
∑ ∑ ��2�\/ª�²ª\ #�&, j� (5.15)
6. Maximum Probability MAX\.ª #�&, j� (5.16)
Table 5.5: GLCM texture descriptors from the MATLAB Image Processing Toolbox
No. Texture Descriptor Formula Equation
No.
1. Correlation (MATLAB)
∑ �\/®}��ª/®����\,ª�¯}¯�\,ª
where #�&, j� is the �&, j�th entry in a normalized GLCM. The standard deviations for the rows and columns of the matrix are: ±] = ∑ ∑ �& − °]�� ∙ª\ #�&, j� ±] = ∑ ∑ �j − °z�� ∙ª\ #�&, j�
where °] and °z are the means for the rows and columns of the matrix respectively
(5.17)
2. Homogeneity (MATLAB) ∑ ��\,ª��2|\/ª|\,ª (5.18)
216
The GLCM texture feature descriptors shown in Table 5.3 (Clausi, 2002) and Table
5.4 (Soh & Tsatsoulis, 1999) are used for texture feature extraction, since these
texture descriptors are more recently proposed in the literature and have
indicated promising results in pattern classification problems using textures. The
MATLAB Image Processing Toolbox possesses texture analysis capabilities. The
*b+Pî$##b$#% function in MATLAB calculates the statistics specified from GLCMs,
which has four texture descriptors, namely: contrast, correlation, energy and
homogeneity. It is noticed that the calculation formulae of the Contrast and Energy
texture descriptors in MATLAB Image Processing Toolbox is similar to the ones
proposed by Haralick (1973) as listed in Table 3.4. Thus, the remaining two
texture descriptors, correlation and homogeneity are applied in this research as
shown in Table 5.5.
In total, 24 GLCM texture descriptors are identified in this research, which are
given in Tables 3.4, 5.3, 5.4 and 5.5, corresponding to equations (3.20) to (3.33)
and equations (5.9) to (5.18) respectively. The GLCM computational parameters:
Number of grey levels, Distance between pixels and Angle (see Section 3.5.3.1.1)
used for texture feature extraction are discussed as follows:
(a) Number of Grey Levels
All GLCM quantization levels i.e., 8, 16, 32, 64, 128 and 256 are used for the
purpose of GLCM texture feature extraction from the grayscale ROIs. This
indicates that for each ROI six different grey levels are computed.
217
(b) Distance between pixels
GLCMs are constructed by identifying neighboring pairs of image cells with
a distance � from each other and incrementing the matrix position
corresponding to the grey level intensity of both cells as indicated in
Section 3.5.3.1.1. In this research, the value of � is chosen as � = 1, in order
to represent the distance between the pixel of interest and the neighboring
pixels in each ROI.
(c) Angle
In GLCMs it is necessary to define the direction of the pair of pixels. The
most common GLCM directions for a given distance � are: ��0°, ��,��45°, ��, ��90°, ��, ��135°, �� as indicated in equations (3.15) to (3.18).
These four directions, i.e., 0°, 45°, 90°, 135° and their symmetric
equivalents, -180°, -135°, 90° and -45° are valid GLCM directions. In this
research, all eight GLCM directions, i.e. 0°, 45°, 90°, 135°, -180°, -135°, 90°
and -45° are computed for each ROI.
The total number of GLCM texture features computed using the GLCM parameters
mentioned above, are reported in Table 5.6. For each ROI, 1152 feature values are
computed using the 24 texture descriptors. This means, for each GLCM texture
descriptor, 48 feature values are calculated. The GLCM texture features calculated
for all the malignant ROIs are shown in Appendix A, Figure A.1, where rows
represent each ROI sample and columns indicate the 1152 texture features for
each ROI sample.
218
Table 5.6: GLCM texture features calculated for each ROI sample
No. GLCM Parameters No. of Texture
Features
1. Texture Descriptors (Table 3.4, 5.3, 5.4 and 5.5) 24
2. GLCM directions (0°, 45°, 90°, 135°, -180°, -135°, 90° and -45°) 8
3. GLCM distance between pixels (� = 1) 1
4. Number of GLCM grey (quantization) levels: 8, 16, 24, 64, 128, 256
6
Texture feature values calculated for each ROI: 1152
5.4.3 Texture Feature Selection
Feature selection needs to be performed in order to select the optimal subset of
features from the 1152 features values obtained in Table 5.6 for each ROI. The
optimal subset of selected features will be used to model a SVM classification
engine for the purpose of pattern classification. As discussed in Section 4.4 of this
thesis, for the purpose of feature selection, a Recursive Feature Elimination (RFE)
technique is applied, namely SVM-RFE. The SVM-RFE technique discussed in
Section 4.4.2 uses F-scores together with the Random Forest (RF) and the SVM.
This feature selection technique is referred to as “F-score + RF + SVM” (Chen & Lin,
2006). In this technique, F-scores are used for the purpose of ranking features
whereas the RF eliminates unimportant features. This algorithm is implemented in
python script, named as fselect.py (Feature selection tool for LIBSVM in Python,
2010) by Chen & Lin (2006).
The fselect.py tool is evaluated with the 1152 texture features computed from all
malignant and benign ROIs (232 samples). Malignant and benign samples
represent two different classes of samples. Firstly, the F-score technique
presented in Section 4.4.1 is used for ranking all 1152 feature values using
equation (4.54). The results obtained for ranking the 1152 features are shown in
219
Appendix A in Figure A.2. The maximum F-score value computed is 2.6425
whereas the minimum F-score value computed is 0.000415 (considered as 0). The
larger the F-score value is, the more the likelihood of that feature being more
discriminative.
Table 5.7: GLCM texture descriptors used to select the optimum subset
of 1056 features
No. GLCM Texture Descriptor Reference Equation
No.
Table
No.
1. Autocorrelation (Soh & Tsatsoulis, 1999) (5.11) 5.4
2. Contrast (Haralick, 1973) (3.21) 3.4
3. Correlation (MATLAB) - (5.17) 5.5
4. Correlation (Haralick) (Haralick, 1973) (3.22) 3.4
5. Cluster Prominence (Soh & Tsatsoulis, 1999) (5.12) 5.4
6. Cluster Shade (Soh & Tsatsoulis, 1999) (5.13) 5.4
7. Dissimilarity (Soh & Tsatsoulis, 1999) (5.14) 5.4
8. Angular Second Moment:
Energy (Haralick, 1973) (3.20) 3.4
9. Entropy (Haralick, 1973) (3.28) 3.4
10. Homogeneity (MATLAB) - (5.18) 5.5
11. Homogeneity (Soh & Tsatsoulis) (Soh & Tsatsoulis, 1999) (5.15) 5.4
12. Maximum Probability (Soh & Tsatsoulis, 1999) (5.16) 5.4
13. Sum of Squares: Variance (Haralick, 1973) (3.23) 3.4
14. Sum Average (Haralick, 1973) (3.25) 3.4
15. Sum Variance (Haralick, 1973) (3.26) 3.4
16. Sum Entropy (Haralick, 1973) (3.27) 3.4
17. Difference Variance (Haralick, 1973) (3.29) 3.4
18. Difference Entropy (Haralick, 1973) (3.30) 3.4
19. Information Measure of
Coefficient 1 (Haralick, 1973) (3.31) 3.4
20. Information Measure of
Coefficient 2 (Haralick, 1973) (3.32) 3.4
21. Inverse Difference Normalized (Clausi, 2002) (5.9) 5.3
22. Inverse Difference Moment
Normalized (Clausi, 2002) (5.10) 5.3
220
Using these F-scores of the 1152 features, the technique presented in Step (2) of
Section 4.4.2 (Chen & Lin, 2006), performs 10-fold CV using the RF technique to
filter out unimportant features. The result of this process is an optimal subset of
features. Applying this technique, an optimal subset of 1056 texture features is
obtained, as shown in Appendix A in Figure A.3. As the optimum subset of features
consists of 1056 texture feature values, this indicates the feature selection process
recursively eliminates 96 feature values. The 96 feature values filtered correspond
to two GLCM texture descriptors, namely the Inverse Difference Moment (in
equation (3.24)) and the Maximum Correlation Coefficient (in equation (3.33)).
Thus, the optimum subset of 1056 texture features shown in Appendix A in Figure
A.3 is obtained using the 22 GLCM texture descriptors shown in Table 5.7.
5.4.4 Feature Normalization
Prior to development of the SVM classification engine, the optimal subset of 1056
texture features obtained in Section 5.4.3, need to be represented in a normalized
scale. In order for the feature data to fit the SVM properly, all 1056 features are
scaled (normalized) in the range between 0 and 1. Feature normalization is
performed using the following expression:
U�O� = +�]�/�8��+�]���@,�+�]��/�8��+�]�� (5.19)
where Ê�O� for O = 1,2,3,… ,1056 represents the feature of interest and min�Ê�O�� and max�Ê�O�� represent the minimum and maximum values corresponding to the
feature of interest �O�. The feature data in Appendix A in Figure A.4 is normalized
in the range between 0 and 1 as shown in Figure A.5.
221
5.5 Classification Engine Development
5.5.1 Feature Labeling and Adjustment
In order for the feature data for SVM training and testing the data needs to be
adjusted in and presented in a proper format to LIBSVM (Chang & Lin, 2010). In
order to serve this purpose, all normalized features values are labeled, where
labels are represented using integers. Normalized feature values corresponding to
their labels are mathematically represented by the matrix , in the form:
=QRRRRS���: O�� ⋯ ��_: O�_ ⋯ ��û: O�û���: O�� ⋯ ��_: O�_ ⋯ ��û: O�û⋮ ⋮ ⋮ ⋮ ⋮���: O�� ⋯ ��_: O�_ ⋯ ��û: O�û⋮ ⋮ ⋮ ⋮ ⋮���: O�� ⋯ ��_ : O�_ ⋯ ��û: O�û XY
YYYZ (5.20)
where � is an integer representing the feature label for � = l1,2,… ,1056t, O
represents the normalized feature value, î indicates the number of texture features
for î = l1,2,… ,1056t and b indicates the number of ROI samples for b =l1,2,… ,232t. During feature labeling the samples belonging to the malignant and benign classes
are labeled using the information from the GT data (in Section 5.2.1). Malignant
samples are represented as the + !î�+%% or î�+%%―1 and benign samples are
represented as the − !î�+%% or î�+%%―2. The LIBSVM feature file obtained after
associating the class memberships of all the samples is shown as Appendix A in
Figure A.6., where the first column represents the class label for each sample
(row).
222
5.5.2 Training and Testing Data Separation
To implement the Non-linear SVM (in Section 4.2.4), the normalized feature needs
to be separated into two distinct sets, i.e. the training set and the testing/validation
set. As observed from Table 5.2, the total number of ROI samples obtained from
the acquired mammography data is 232, out of which 118 are malignant samples
�î�+%%―1� and the remaining 114 are benign samples �î�+%%―2�. In order to split the feature data into the training and testing sets, the Holdout
method presented in Section 4.1.3.1 is adopted, where two-third (70 percent) of
the samples from both classes are allocated to the training set and the remaining
one-third (30 percent) of the samples from both classes are allocated to the testing
set. The specification of the samples in training and testing sets is indicated in
Table 5.8.
Table 5.8: Ratio of samples used for training and testing from the UMMC
and MIAS datasets
Class Number of
Samples
Training Set
(70% Samples)
Testing Set
(30% Samples)
Malignant (+ve) �î�+%%―1� 118 82 36
Benign (-ve) �î�+%%―2� 114 80 34
Total Samples: 232 162 70
5.5.3 SVM Model Development
The SVM is implemented in this research using LIBSVM library (Chang & Lin,
2010) integrated into MATLAB. As observed from Table 5.8, a total of 162 samples
from both classes are used for SVM training (memorization/learning), while the
remaining 70 samples are used for testing the accuracy of the trained model or
SVM classification engine for unseen data samples. Since the training accuracy of
223
the SVM classifier is evaluated using the 10-fold CV approach discussed in Section
4.1.3.2 of this thesis, thus, on each CV fold, training and testing samples are
selected randomly, so as to ensure that the developed classification engine does
not overfit the training data. The SVM training engine proposed for constructing
the classification engine and performing SVM hyperparameter optimization is
illustrated in Figure 5.27.
5.5.3.1 SVM Parameter Optimization
The training or learning accuracy of the SVM classification engine is estimated by
tuning the error penalty parameter, � (in equation (4.29)). In this research, the
RBF (Gaussian) kernel (in equation (4.44)) is used with Non-linear SVM as
discussed earlier in Section 4.2.4. The parameter � in the RBF kernel which
controls the width of the Gaussian needs to be optimized with respect to the SVM
hyperparameter �. Thus, two SVM hyperparameters ��, �� need to be determined
in order to construct a classifier with an optimum balance between its
memorization and generalization capability.
The Grid Search method proposed by Hsu et al. (2003) and discussed in Section
4.2.4 is used in this research for SVM hyperparameter optimization, as indicated in
Figure 5.27. In the Grid Search method, exponentially growing sequences of
parameters ��, �� are used to identify optimum parameter values with respect to
the best 10-fold CV accuracy. In this trial and error procedure, sequences of
parameters in the range, � = 2�, 2�… , 2�� and � = 2/� , 2/�., … , 2�� are
evaluated for 100 × 100 = 10,000 combinations respectively.
224
Perform 10-fold CV using 70% data for
Training and 30% for Testing
Reselect Optimal SVM Hyperplane
Parameters using Grid Search
Trained Model
(SVM Classification Engine)
10-fold CV Accuracy for 100
trials
End of SVM Training
Training Data (162 Samples)
Selected randomly
Bad Good
Start of SVM
Training
SVM Parameter
Optimization
Initialize SVM Hyperplane
Parameters using Grid Search
Figure 5.27: The SVM training engine proposed for constructing the classification
engine and performing hyperparameter optimization
225
For each pair of ��, �� the 10-fold CV performance of the trained model is
measured by splitting the training data (162 samples) into two smaller CV subsets,
using the Holdout method presented in Section 4.1.3.1. Using this method, the CV
training subset contains two-thirds of the training samples (113 samples), whereas
the CV testing subset contains the remaining one-third of the training samples (49
samples). The selection of the CV training and CV testing subsets is repeated 100
times for 10-fold CV trials, where on each trial, the samples used for CV training
and CV testing are selected randomly.
Iterating different parameter combinations of ��, �� experimentally, the optimum
SVM parameters using Grid Search are found to be: � = 64 and � = 0.001953125,
which obtain the highest 10-fold CV accuracy of 87.83 percent as indicated in
Figure 5.28. The training (memorization) accuracy1 of the SVM classification
engine is calculated using the following expression:
�b+&)&)*Mîîíb+îP = �/�0 × 100% (5.21)
where �' represents the total number of samples correctly classified by the SVM
and �2 represents the total number of samples used for CV testing. As 49 samples
are used for CV testing ��� = 49�, the 10-fold CV results obtained indicate that 48
out of 49 samples are classified correctly by the SVM, thus, �' = 48. Using equation
(5.21) with parameter values �' = 48 and �� = 49 a training accuracy of 97.6
percent is achieved for the SVM classification engine. This indicates that the
developed classification engine has good learning capability.
1Training accuracy is the measure of the memorization and learning capability of the classifier. Training accuracy is calculated in percentage of the total number of samples used for SVM testing divided by the total number of samples correctly classified by the SVM.
226
Figure 5.28: Grid Search for selection of optimal of SVM hyperparameters ��, �� 5.5.3.2 Probability Estimation
Besides conducting classifications, SVMs also compute the probabilities for each
class (Wu et al., 2004). This supports the analytic concept of generalization and
certainty. Given that b\ª is an estimate for the probability of the output of pairwise
classifiers between class & and class j (i.e., b\ª 3 P�P = l&, jt, O�, b\ª + bª\ = 1) and
that #\ is the probability of the &cd class, the probability # = �#�, … , #_� of a class
can be derived via a quadratic programming (QP) problem (Oskoei & Hu, 2008):
min� ∑ ∑ �b\ª #ª − bª\ #\�� , ∑ #\ = 1#\ ≥ 0, ∀&_\s�ª:ª5�_\s� (5.22)
The pairwise probability information defined in equation (5.22) is computed using
the LIBSVM library (Chang & Lin, 2010), to estimate the probabilities of the tested
227
samples. The probability estimates (decision values) for the testing set, provides
additional information for selection of samples with higher confidence measures
(probabilities), as will be discussed in Section 5.5.3.5 later.
5.5.3.3 SVM Training
SVM training is performed by integrating the LIBSVM library (Chang & Lin, 2010)
into MATLAB. After obtaining the optimal pair of SVM hyperparameters ��, �� (see
Section 5.5.3.1), the SVM is trained using the 162 training samples as indicated in
Table 5.8 and Figure 5.27. The LIBSVM MATLAB executable, svmtrain.mexw32 is
employed for SVM training as shown in Appendix B in Figure B.1.
The optimized SVM hyperparameters ��, �� obtained from the Grid Search method
in Section 5.5.3.1 are used to model a Non-linear SVM for binary classification, as
shown in Appendix B in Figure B.2. In Figure B.2, samples.txt represents the 162
training samples in Table 5.8. The ‘-v 10’ parameter used in the training string
performs 10-fold CV, where the result obtained is the memorization accuracy of
the trained classifier. During the SVM training, the SMO process finishes on the
5494th iteration obtaining the highest 10-fold CV accuracy of 87.83 percent as
indicated in Section 5.5.3.1 and Figure 5.28.
The trained model generated in MATLAB after SVM training is shown in Appendix
B in Figure B.3, where the SVM model parameters are shown in Figure 5.29. As
observed from Figure 5.29, the trained model contains 52 support vectors (SVs),
defined by the constraint 0 ≤ â\ ≤ � in equation (4.34), where the malignant class
�î�+%%―1� has 29 SVs and the benign class �î�+%%―2� has 23 SVs. The total number
of bounded SVs (BSVs) computed by the model is 10, justifying that the condition
228
â\ = � (in equation (4.35)) is true. The ba$ parameter defined as ba$ = −� in the
decision function in equation (4.42) is computed to be: ba$ = 13.3317.
Figure 5.29: SVM classification engine ‒ Trained model
The separating boundaries between the two classes (malignant and benign) of the
training data in the SVM classification engine (in Figure 5.29) are shown in Figure
5.30. This plot is drawn using the “SVM Toy” tool, svm-toy.exe included in the
LIBSVM (Chang & Lin, 2010) library. As observed from Figure 5.30, the purple
(dark) dots represent malignant samples, while the green (light) dots represent
benign training samples. A training sample in the purple region indicates that
sample is classified as malignant �î�+%%―1�, whereas a sample in the green region
indicates that sample is classified as benign �î�+%%―1�. Observing the non-linear
soft-margin boundaries between the two classes of training data, it is suitable to
imply that samples in both classes are well separated (with a only few
misclassifications), which indicates that the trained model has good learning and
memorization capability,
229
Figure 5.30: Separating boundaries of the SVM classification engine in Figure 5.29
5.5.3.4 SVM Testing and Validation
During the testing and validation phase, the classification engine developed in
Section 5.5.3.3 is used with the remaining 30 percent testing samples (70 samples)
as shown in Table 5.8 to test the accuracy of the developed system.
The LIBSVM MATLAB executable for SVM testing, svmpredict.mexw32 is evaluated
for testing and validating the classification engine, as shown in Appendix B in
Figure B.4. The samples used for SVM testing are in the exact same format as the
training data samples shown in Appendix A in Figure A.6. The class membership
230
associated with the SVM testing samples can be taken as either class, malignant
�î�+%%―1� and/or benign �î�+%%―2�. This is because the class labels are only useful
for computing the �–fold CV accuracy during SVM training, and since the labels of
the testing samples in are unknown, they can be considered as either class. In this
research all SVM testing samples are represented as benign �î�+%%―2�. So, during
testing the class labels are ignored as they going to be predicted by the SVM
classification engine.
SVM testing and validation is performed in MATLAB using the SVM classification
engine and the testing samples as shown in Appendix B in Figure B.5. The ‘-b 1’
parameter used in the SVM prediction string computes the probability estimates of
the tested samples, as discussed in Section 5.5.3.5. The classification accuracy of
the SVM for testing the 70 samples (in Table 5.8) is found to be 97.14 percent,
which indicates that out of 70 samples 68 samples have been correctly classified
by the SVM classification engine.
The resulting parameters ‘outputlabel’ and ‘probability’ in Figure B.5 contain the
classification results as the ‘predicted class labels’ and ‘probability estimates’
respectively for the tested samples, as shown in Figure 5.31. As observed from
Figure 5.31, the first column represents the predicted class labels, i.e. î�+%%―1 or
î�+%%―2 for the tested samples, whereas the second and third columns represent
the probability of î�+%%―1 �#ûh0���� and î�+%%―2 �#ûh0���� for each testing sample.
5.5.3.5 Logic System for False Positive (FP) Reduction
The SVM testing and validation results as shown in Figure 5.31 include probability
estimates of the tested samples. These probability estimates are used to model a
231
decision-logic system, to reduce the number of FPs; serving as the one of the key
objectives outlined in this research in Sections 1.2 and 1.3.
Figure 5.31: SVM testing and classification results using LIBSVM in MATLAB
As shown in Figure 5.31, each row represents a testing sample, where the first
column is the predicted label of the tested sample. The second and third columns
232
represent the probability estimates of the malignant samples �+ !î�+%%� and
benign samples �− !î�+%%� respectively. The probability estimates indicate the
belongingness of a testing sample to a class. For a testing sample, if the probability
of malignant is higher than benign, then that sample is classified as malignant
while if the probability of benign is higher than malignant then that sample is
classified as benign. This indicates that the class with the higher probability will be
predicted class label. The probability estimates for each sample are computed
such that:
#ûh0��� + #ûh0��� = 1 (5.23)
As observed from equation (5.23), summing the malignancy and benign
probability values #ûh0��� and #ûh0��� for any testing sample, equals to 1. In order to
model a decision-logic system to reduce the number of FPs, the differences
between the probabilities for both classes is calculated for each testing sample,
using the following expression:
� = |#ûh0��� − #ûh0���| (5.24)
The probability differences (computer using equation (5.24)) of the misclassified
samples are compared with the probability differences of the correctly classified
samples in order to determine a threshold, which is used to reduce the FPF.
Inspection of the probability difference data from the 70 testing samples shows
that a threshold value of ' = 0.08 can reduce the FPs. A pseudo code of the
decision-logic system using equation (5.24) in MATLAB is given in Figure 5.32. The
experimental results obtained from testing the decision logic-system are
presented and discussed in Section 6.1.4.1.3.
233
Figure 5.32: Decision-logic system for reduction of false positives (FPs)
5.6 Summary
This chapter presented the modeling of the framework (system) proposed in
Chapter 1 for the classification of benign and malignant abnormalities in digital
mammograms. As discussed in Chapter 3, the proposed framework is composed of
two main techniques, namely, image processing and machine learning. The image
process and machine learning techniques identified in Section 3.3, are applied in
the proposed system in this chapter, which are discussed in Section 5.1 and 5.2.
The modeling of the system consists of three main stages, namely: Mammogram
Image Processing (Section 5.3), Texture Feature Extraction and Selection (Section
5.4) and Classification Engine (Section 5.5). Sections 5.3 through 5.5 describe each
stage in detail during the development of the proposed framework.
The preliminary testing results of the developed framework in Section 5.5.3.4 and
Section 5.5.3.5 show promising results for the classification of malignant and
benign abnormalities in digital mammograms. The framework developed in this
chapter is tested thoroughly, where the experimental results are presented and
discussed in Chapter 6.
234
CHAPTER 6
EXPERIMENTAL RESULTS AND DISCUSSION
6.0 Overview
This chapter presents the experimental results of the developed system in Chapter
5. Section 6.1 presents and discusses the SVM training results relative to the
memorization and learning of the binary SVM classifier. Section 6.1 also presents
and discusses the SVM testing and validation results for unseen samples. In order
to perform a comparative research, Section 6.2 presents the experimental results
obtained after evaluating the developed framework using different machine
learning algorithms other than the SVM. The experimental results of the compared
machine learning models are discussed in the last part of Chapter 6.
6.1 Experimental Results of Proposed Framework
6.1.1 Image Segmentation Performance Indices
The accuracy of the mammogram segmentation (Stage 1 in Figure 5.2) algorithm
in this research is evaluated by deriving quantitative measures by comparing each
segmented mammogram mask with its corresponding gold standard. In this
research, the gold standard is obtained by manually segmenting the breast region
from the background region for all mammogram images acquired. To serve this
purpose, the boundary of the breast is traced to extract the real breast region,
which results in a Ground Truth (GT) image as shown in Figure 6.1.
235
Quantitative measures using the Receiver Operating Characteristics (ROCs) (see
Section 3.3.6) are used to describe the accuracy of the mammogram segmentation
process. The region (mask) obtained from the segmentation result which matches
the GT image, is denoted as the True Positive (TP) pixels, which expresses that the
segmentation algorithm has found a portion of the breast. The pixels shown in the
GT image but not shown in the mask are denoted as False Negative (FN) pixels,
which are the missing pixels in the breast region. Finally, the pixels not in the GT
image, but in the shown in the mask, are denoted as False Positive (FP) pixels.
Figure 6.1: Image segmentation performance indices: TP, FP and FN
236
Using the mammogram segmentation performance indices (TP, FP and FN) in
Figure 6.1, two metrics relating to the segmentation performance are derived,
namely: Completeness (CM) and Correctness (CR). In mammogram segmentation,
CM is the percentage of the GT region, which describes the segmented region,
using the following expression:
�$k#�!'!)!%%�CM� = ?4�?427K� (6.1)
CM ranges from 0 to 1, with 0 indicating that none of the regions are properly
partitioned, and 1 indicating that all the regions were segmented. For example, a
value of CM = 0.92 indicates a 92 percent overlap with the GT image. Similarly, CR
represents the percentage of correctly segmented breast region (profile), using the
following expression:
�$bb!î')!%%�CR� = ?4�?4274� (6.2)
Similar to CM, the optimum value for CR is 1 and the minimum value is 0. Lower
values of CM indicate over-segmentation, whereby a region in the GT is
represented by two or more regions in the examined segmented image. Similarly,
under-segmentation is defined for CR, where two or more regions in the GT are
represented by a single region in the segmented image. In mammogram
segmentation, the segmentation algorithm is considered accurate if the percentage
of CM and CR is greater than 95 percent. A more general measure of the
mammogram segmentation performance is achieved by combining CM and CR into
a single measure known as Quality (Q), using the following expression:
237
Ãí+�&'P�Q� = ?4�7K2742?4� (6.3)
Similarly, the optimum value for Q is 1 and the minimum value is 0. Results
obtained from mammogram preprocessing (in Figures 5.1 and 5.2) indicate some
influence on the effectiveness of the segmentation algorithm, but since the noise
removal and background/artifact suppression algorithms are not image
enhancement algorithms, there are no GT images present. Thus, it is considered
non-trivial to quantitatively measure the effective of the mammogram
preprocessing.
6.1.1.1 Image Segmentation Results
The mammogram segmentation algorithm (in Figures 5.1 and 5.2) is evaluated on
all the 582 mammogram samples as shown in Table 5.1. To demonstrate the
robustness of the segmentation algorithm, it has been evaluated on mammograms
with differing breast densities such as fatty, fatty fibroglandular and dense
fibroglandular tissues.
Segmenting all the 582 mammogram images (in Table 5.1), the average CM and CR
obtained are 0.996 and 0.981 respectively, signifying that the mammogram
segmentation algorithm is robust with respect to different tissue densities. This
implies that the average proportion of the segmented breast region detected by
the algorithm is 99.6 percent, while 1.9 percent of the background is mislabeled as
the breast region.
With few exceptions the mammogram segmentation algorithm performed well
with sufficient reliability to retain the nipple in the breast region (profile). After
238
segmentation is completed, it is computed that the average breast region contains
approximately 208,400 pixels. Thus, on average, each segmented mask misses 366
pixels from the breast region and mistakes 1742 pixels from the background
region as breast pixels, which gives a quality of Q = 0.98. The adaptability of the
segmentation algorithm in terms of tissue density is illustrated in the three
following experiments.
Experiment 1 ― Fatty Tissue
The first experiment deals with mammograms which predominantly comprise of
fatty tissue. The segmented image closely approximate to the breast region as
represented by the GT image. The quantitative measures indicate that the
segmented breast regions are marginally under-segmented (2 percent), but do
contain the breast region in their entirety. The mean CM and CR values for all fatty
tissues are computed to be 0.99 and 0.96 respectively.
Experiment 2 ― Fatty-Fibroglandular Tissue
The second experiment deals with mammograms which comprise of fatty
fibroglandular tissue. The mean CM and CR values for all fatty fibroglandular
tissues are computed to be 1.00 and 0.99 respectively. Since, the CM and CR values
are closest to the optimum values that can be obtained, this indicates that the
segmentation error is very small, i.e., less than 1 percent.
Experiment 3 ― Dense-Fibroglandular Tissue
The final experiment deals with mammograms comprising of dense fibroglandular
tissue. The mean CM and CR values for all dense fibroglandular tissues are
computed to be 1.00 and 0.98 respectively. For dense fibroglandular tissues the
239
nipple in the breast profile in all the mammograms is retained and the segmented
breast region compares well with the GT images.
As a conclusion to the three experiments above, the mammogram segmentation
algorithm is invariant to changes in tissue density. However, no segmentation
algorithm can be considered 100 percent robust, this is due to the heterogeneous
nature of mammograms. Typical mammogram acquisition problems include:
scanner induced artifacts, excessive background noise and scratches and dust
which influence the reliability of the segmentation algorithm. The mammogram
segmentation results indicate that, for all 582 mammograms evaluated, only 9
mammograms (1.5 percent) fell marginally short of the 95 percent accuracy
indicator and 3 mammograms (0.5 percent) were over-segmented, which
attributed primarily to indistinct boundaries.
The reason for the 2 percent segmentation inaccuracy is because these 12
mammograms are considered as special cases, which have a highly non-uniform
background and less contrast in the area above the breast tissue region. So, the
segmentation results in a roughly extracted breast contour corresponding to the
breast region. The CM and CR measures of the 9 (1.5 percent) oversegmented
mammogram images are computed to be 0.87 and 0.99 respectively, which
indicates that the GT image contains the entire segmented region.
6.1.2 Feature Selection Results
Texture features are computed using the GLCMs of all ROI samples (malignant and
benign) for the purpose of binary classification using SVM. The feature selection
algorithm evaluated in this research is known as “F-score + RF + SVM” technique
240
(Chen & Lin, 2006), which is discussed in detail and the experiments are reported
in Section 4.4 and Section 5.4.3 of this thesis respectively.
Initially, 24 GLCM texture descriptors are used for the purpose of feature selection,
as indicated in Table 5.6. After feature selection (Chen & Lin, 2006), the optimum
subset of texture features is computed to be 1056, which corresponds to 22 GLCM
texture descriptors as indicated in Table 5.7. This indicates that the Recursive
Feature Elimination (RFE) technique eliminates 2 GLCM texture descriptors
corresponding to 96 texture feature values. The optimum subset of 1056 texture
features obtains the highest 10-fold CV accuracy of 82.30 percent. The following
section discusses the feature selection results obtained using the proposed
technique.
6.1.2.1 Discussion of F-score Results
The F-score feature ranking algorithm (in Section 4.4.1) uses a RFE technique,
namely the SVM-RFE as discussed in Section 4.4.2. In order to compute the F-
scores for the GLCM texture features using the SVM-RFE based approach, binary
confusion matrix performance indices (TP, FP and FN) need to be computed at
first, as indicated in Table 3.1. Prior to computation of F-scores, the �b!î&%&$) and
!î+�� for each texture feature is computed using the following expressions:
�b!î&%&$) = ?4?4274 (6.4)
!î+�� = ?4?427K (6.5)
241
Using the precision and recall values obtained from equations (6.4) and (6.5), F-
scores for each texture feature are calculated using the following expression:
Ê − %î$b! = �×1�û\�\��×9û0hh1�û\�\��29û0hh (6.6)
The F-scores computed for the GLCM texture features are shown in Appendix A in
Figure A.2, whereby the optimum subset of features selected using the proposed
technique is shown in Appendix A in Figure A.3. Since the proposed feature
selection algorithm in this research obtains a 10-fold CV accuracy of 82.30 percent
using the “F-score + RF + SVM” technique, this indicates that the optimum subset
of features selected has a negligible correlation between each feature. This is
because during SVM-RFE, features with lower F-scores are eliminated.
6.1.3 SVM Training Validation
After obtaining the optimal pair of SVM hyperparameters ��, �� (in Section 5.5.4.1),
the SVM is trained for binary classification using the 162 training samples as
indicated in Table 5.8.
Using the training approach discussed in Section 5.5.3.3, the highest 10-fold CV
accuracy of the SVM classification engine obtained is 87.83 percent, as indicated in
Figure 5.28. The training accuracy of the classification engine is calculated using
equation (5.21), which results in a training accuracy of 97.60 percent. The training
accuracy indicates that the developed classification engine has good learning and
memorization capability. The separating boundaries (soft-margin) between the
two classes of the training data, î�+%%―1 (malignant) and î�+%%―2 (benign) is
illustrated in Figure 5.30.
242
Prior to SVM training, optimum SVM hyperplane parameters ��, �� need to be
determined. As mentioned throughout this thesis, for the purpose of SVM
hyperparameter optimization, 10-fold CV is extensively used. During CV all SVM
training samples (in Table 5.8) are trained and validated in order to generalize the
memorization accuracy of the SVM classification engine. The main reason for
conducting 10-fold CV is to ensure that the SVM classification engine does not
overfit the training data.
For the purpose of applying 10-fold CV, the 162 training samples (in Table 5.8) are
split into CV training and CV testing sets such that, 70 percent of the total samples
(113 samples) from each class are used for CV training and the remaining 30
percent samples (49 samples) from each class are used for CV testing. This
iterative procedure is repeated for 100 trials for 10-fold CV, where on each trial
the CV training and CV testing data samples are selected randomly.
The Grid Search method proposed by Hsu et al. (2003) (in Section 4.2.4) is used for
SVM hyperparameter tuning in this research. In the Grid Search method,
exponentially growing sequences of parameters ��, �� are used to identify SVM
hyperparameters obtaining the best 10-fold CV accuracy of 87.83 percent (in
Section 5.5.3.3). After Grid Search is complete, the optimum SVM hyperplane
parameters are found to be: � = 64 and � = 0.001953125, as shown in Figure
5.28. Thus, using the optimum set of SVM hyperparameters obtained from the Grid
Search method, an average SVM training accuracy of 97.60 percent is obtained
using the 49 CV testing samples (see Section 5.5.3.3).
243
6.1.3.1 Discussion of SVM Training Results
In the C-SVM classification model (Hsu et al., 2003) applied in this research, the
parameter � is a SVM hyperparameter that defines the trade-off between the
training error and complexity of the model (classification engine). In the dual
Lagrangian formulation, the parameter � (in equation (4.30)) defines the upper
bound of the Lagrange multipliers â\ , hence, it defines the maximal influence the
sample can exert on the solution.
For the trained model developed in Figure 5.29, the SVM hyperparameter � affects
the training and memorization accuracy of the SVM classification engine. The
reason for this is, since there are 10 bounded SVs (BSVs) in the trained model,
thus, â\ = � (in equation (4.35)). Due to this, the Grid Search technique selects the
parameter � that defines the optimum trade-off between the training error and the
complexity of the model, with parameter � = 64, signifying that the training data
has significant noise. Thus, by using a smaller value of parameter � in the
developed model, the results of the SVM classification mapping are smoother with
a lower noise consideration. The RBF kernel parameter �, in the SVM classification
engine controls the width of the RBF (Gaussian) kernel. The � parameter is related
to ±, which is defined by the following expression:
� = ��¯² (6.7)
where ±� is the variance of the resulting Gaussian hypersphere. The optimum
value of the SVM hyperparameter � in equation (6.7) found using Grid Search is
computed to be: � = 0.001953125. So, the value of ± can be calculated using the
following expression:
244
± = :� ��ô� (6.8)
where ± is computed to be ± = 16 using � = 0.001953125. The value of ± for the
trained classifier is acceptable, since any value of ± below 0.01 is considered small
and any value of ± above 100 is considered large. The reason for this is, as the
parameter ± acts as an important hyperparameter during SVM training, small
values of ± lead the model close to overfitting the training data, while large values
of ± tend to over-smooth the training data. From the statistical learning theory
point of view, small ± values lead to a higher VC-dimension, meaning that too many
features are used for machine learning which leads to overfitting, while large ±
values lead to a lower VC-dimension, signifying that too few features are used to
model the classification engine. Thus, the value ± = 16 is acceptable to model the
SVM classification engine using the RBF kernel.
6.1.4 SVM Testing and Validation
The accuracy of SVM testing and validation of is a gauge to evaluate the capability
of the developed framework, namely the capability to classify between malignant
and benign samples. In this research SVM testing and validation is performed by
integrating the LIBSVM v3.0 library (Chang & Lin, 2010) into MATLAB as indicated
in Section 5.5.3.4.
The trained model in Figure 5.29 is validated with the 70 testing samples (in Table
5.8) in order to classify previously unseen (untrained) samples. As observed from
Figure B.5 in Appendix B, the SVM testing accuracy obtained for an average of 100
trials using 70 testing samples (selected randomly on each trial) is found to be
245
97.14 percent. In addition, the SVM probability estimates of the tested samples
(see Section 5.5.3.4) are obtained with the SVM classification results (class labels).
The probability estimates (or scores) can be taken as a measure of confidence
during classification of the testing samples, as indicated in Figure 5.31 and Figure
5.32. The experiments performed in this research are presented in Section 6.1.4.1
and discussed in Section 6.1.4.2.
6.1.4.1 SVM Classification Results
The framework developed in this research for the classification of malignant and
benign abnormalities (in Figure 5.2) is tested using a Dell XPS 430 Workstation, with
a 3.00 GHz Intel Core2 Quad Processor and 8.00 GB of RAM. The time taken for
testing one sample approximately takes 4 seconds, which varies based on the
configuration of the computer used and the number of samples tested. The
following sections present the experiments performed in order to meet the
objectives and contributions of this research outlined in Section 1.2 and Section
1.3 respectively.
6.1.4.1.1 Optimum ROI Size Selection
In general it is difficult to determine the size of the neighbourhood or the Region of
Interest (ROI) that should be used to extract the relevant GLCM texture features
from the abnormal regions (mass lesions and MCCs). If the size of the ROI is too
large, small lesions may be missed; while if the ROI size is too small, parts of large
lesions may be missed.
The primary contribution of this research as indicated in Section 1.3 is to
determine the most suitable ROI (neighbourhood) size in order to perform
246
optimum texture feature extraction. This specifically addresses the problem of
predeterming the ROI size for feature extraction. Thus, in this research, seven
common ROI sizes have been evaluated as discussed in Section 5.4.1, namely: 48 ×
48 pixels, 64 × 64 pixels, 96 × 96 pixels, 110 × 110 pixels, 128 × 128 pixels, 136 ×
136 pixels and 148 × 148 pixels.
Table 6.1: Comparison of classification accuracy using different ROI sizes
No. ROI Size
(in pixels)
Optimum SVM
Hyperparameters
Average SVM Accuracy
for 100 trials
1. 48 × 48 C = 64, γ = 0.0078125 86.56%
2. 64 × 64 C = 1024, γ = 0.001953125 89.33%
3. 96 × 96 C = 256, γ = 0.001953125 93.87%
4. 110 × 110 C = 32, γ = 0.00390625 94.58%
5. 128 × 128 C = 64, γ = 0.001953125 97.60%
6. 136 × 136 C = 512, γ = 0.0009765625 93.53%
7. 148 × 148 C = 256, γ = 0.00390625 92.76%
Testing the significance of the ROI sizes is performed using 70 testing samples (30
percent of the total ROI samples) with the developed SVM classification engine.
The experimental results obtained using different ROI sizes and their tuned SVM
hyperparameters, are shown in Table 6.1. As indicated from Table 6.1, the ROI size
of 128 × 128 pixels obtains the highest performance of 96.60% in terms of
classification between malignant and benign ROIs. Further testing for significance
shows that using a ROI size of 128 × 128 pixels results in the lowest number of FPs
and FNs (see Table 6.1) as compared to the other six ROI sizes.
In addition, performing analysis on the GT data, the minimum and maximum
diameter in pixels of a circle enclosing all malignant and benign abnormalities is
247
found to be 48 and 130 pixels respectively. Given the above reasons, it is confirmed
that a 128 x 128 pixel square ROI (or a 128 pixel circle diameter) is a near
optimum to the value that can be used to extract all the abnormal (malignant and
benign) regions. All experiments performed from here onwards use a ROI size of
128 × 128 pixels for the purpose of texture feature extraction.
6.1.4.1.2 SVM Testing
In this research, the SVM classification engine is developed using 162 training
samples (70 percent of the total ROI samples) for a binary classification problem,
where malignant samples are taken as the #$%&'& !�+ !� class and benign
samples are taken as the )!*+'& !�− !� class. Thus, representing the ROI
samples as positive and negative instances of a binary classification problem, a
confusion matrix can be derived, as indicated in Figure 6.2.
Figure 6.2: Binary classification confusion matrix
Testing the 70 samples indicated in Table 5.8 with the SVM classification engine in
Figure 5.29, the resulting confusion matrix obtained with performance indices TP,
FP, FN and TN is shown in Figure 6.3.
True Positive (TP)
False Positive
(FP)
False Negative
(FN) True Negative
(TN)
p n Total
Actual value
Prediction outcome
P’
n’
P’
N’
Total P N
248
Figure 6.3: Confusion matrix after SVM testing
(malignant is the + ! class and benign is the − ! class)
The SVM testing results obtained in the binary confusion matrix in Figure 6.3 show
that, 30 out of the total 31 malignant �+ !î�+%%� samples are classified correctly
by the SVM, whereas 38 out of the total 39 benign �− !î�+%%� samples are
correctly classified by the SVM. This indicates that only one sample in both classes
is misclassified. Thus, in total 68 out of the 70 tested samples (in Table 5.8) are
classified correctly by the SVM, which give a binary classification accuracy of 97.14
percent as indicated in Appendix B in Figure B.5. Using the confusion matrix
results in Figure 6.3, the four binary classification performance metrics defined in
Table 3.2, namely the TP, FP, TN and FN are computed as shown in Table 6.2. The sensitivity and specificity metrics (in equations (3.1) and (3.2)) from the
confusion matrix performance metrics are computed to be 0.9710 and 0.9706
respectively, where the minimum and the optimum values of both are 0 and 1
respectively. The classification accuracy is computed using equation (3.3), which is
found to be 97.15 percent, where 68 out of the total 70 tested samples are
classified correctly by the SVM. Since the sensitivity, specificity and accuracy
values are greater than 0.95 (95 percent), thus, the performance of the developed
framework is acceptable.
TPs = 35 FPs = 1
FNs = 1 TNs = 33
Total samples
36
34
Positive class
(Malignant)
Negative class
(Benign)
70
249
Table 6.2: Binary classification performance metrics using the SVM as the
learning machine
Binary Classification Performance Metrics
Equation No.
Value
True Positive (TPs) - 35
False Positive (FPs) - 1
False Negatives (FNs) - 1
True Negatives (TNs) - 33
Sensitivity (3.1) 97.10%
Specificity (3.2) 97.06%
Accuracy (3.3) 97.14%
True Positive Fraction (TPF) (3.4) 0.9710
False Positive Fraction (FPF) (3.5) 0.0290
The True Positive Fraction (TPF) (also known as the sensitivity) and False Positive
Fraction (FPF) metrics are calculated using equations (3.4) and (3.5), which are
found to be 0.9710 and 0.0290 respectively as shown in Table 6.2. The TPF
determines the performance of the SVM classification engine on identifying
positive (malignant) samples correctly from all positive samples tested. In
contrast, the FPF determines how many incorrect positive results occur among all
negative (benign) samples tested.
To visualize binary classification results of the developed framework in Figure 5.1,
an ROC curve is plotted using the 70 testing samples (in Table 5.8), as shown in
Figure 6.4. Each instance (testing sample) in the binary confusion matrix in Figure
6.3 is represented as one point in the ROC space in Figure 6.4.
250
Figure 6.4: ROC curve of SVM classifier for testing with 70 samples
(malignant is the + ! class and benign is the − ! class)
The Area Under Curve (AUC) for the ROC curve in Figure 6.4 is found to be
MN = 0.97574. The optimum value for the AUC is 1, where ROC curves with
MN ≥ 0.9 are rated as optimum classification results. As observed from the plot in
Figure 6.4, the ROC curve follows close to the left-hand border and then the top
border of the ROC space, which indicates that the developed framework produces
optimum results for classification between malignant and benign samples.
6.1.4.1.3 False Positive (FP) Reduction Results
The decision-logic system presented in Section 5.5.3.5 and shown in Figure 5.32,
reduces the number of FPs for the confusion matrix in Figure 6.3. Each FP instance
satisfying the condition �' < 0.08� in Figure 5.32 is classified as TN instead of FP,
the result of which is shown in the confusion matrix in Figure 6.5.
251
Figure 6.5: Confusion matrix after implementation of decision-logic system
(malignant is the + ! class and benign is the − ! class)
Using the confusion matrix in Figure 6.5, the FPF is calculated using equation (3.5)
and is found to be 0, which is an ideal value for the FPF. Using the proposed
decision-logic system with a small number of testing samples (70 samples in this
case) a FPF of 0 is achievable. However using this decision-logic system, an ideal
FPF of 0 cannot be guaranteed unless a larger amount of samples are tested.
One of the limitations in this research concerns the number of mammogram
samples acquired for development of the computerized breast cancer detection
system. The total number of mammography images obtained from University
Malaya Medical Centre (UMMC) is limited due to the fact that UMMC have only
recently implemented digital mammography in 2008. Thus, over a course of nearly
two years, only a limited number of malignant and benign cases (in Table 5.1) are
available from the UMMC in digital format.
6.1.4.2 Discussion of SVM Classification Results
This section summarizes the SVM classification results obtained from experimental
testing in Section 6.1.4.1, where four major experiments are performed. All results
presented in this section are evaluated on the UMMC and MIAS ROI samples in
Table 5.8.
TPs = 35 FPs = 0
FNs = 2 TNs = 34
Positive class
(Malignant)
Negative class
(Benign)
252
The first experiment presented in Section 6.1.4.1.1 evaluates different ROI sizes in
order to determine an optimum ROI size, the experimental results of which are
presented in Table 6.1. The optimum ROI size is found to be 128 × 128 pixels with
a classification accuracy of 97.60 percent between malignant and benign ROI
samples. Since the classification accuracy is greater than 0.95 (95 percent), thus,
the performance of the proposed model is acceptable.
The second experiment in Section 6.1.4.1.2 computes the four binary classification
performance metrics (TP, TN, FP and FN) using the SVM testing results from the
first experiment. The performance metrics are used to plot an ROC curve obtained
by testing the 70 samples (in Table 5.8), as shown in Figure 6.4. The ROC curve
yields an AUC of MN = 0.97574. Based on a collective comparison of the results
obtained from the first and second experiment, the following observations are
made:
1. All binary performance metrics in Table 6.2 are greater than 95 percent.
2. The FPF in equation (3.5) is less than 5 percent.
3. The ROC curve follows close to the left-hand border and the top border of
the ROC space.
4. The ROC MN ≥ 0.9.
These observations indicate that the developed system can classify between
malignant and benign ROIs with an average classification accuracy of 97 percent.
Since the classification accuracy of the developed system is greater than the
baseline of 95 percent, thus it is confirmed that the developed framework shown in
in Figure 5.1 produces promising classification results.
253
The third experiment in Section 6.1.4.1.3 gives attention on reducing the number
of FPs obtained from the SVM classification results in Table 6.2. The number of FPs
effect the FPF (in equation (3.5)), which can be reduced by applying a decision-
logic system (in Figure 5.32) using the probability estimates of the tested samples
from the SVM classification results. Applying the decision-logic system confirms
that number of FPs and the FPF can be minimized at a low cost. However, since the
number of samples in the MIAS and UMMC datasets is less the accuracy of the FPF
the reduction algorithm cannot be tested in depth.
6.2 Comparison of Proposed Framework with Other Techniques
In order to estimate the performance of the SVM based model, different machine
learning algorithms other than SVM are evaluated. Since, ANNs have similar
structure to that of SVMs, thus, they are used in this research comparison with the
proposed SVM framework. Traditional and modern ANN based machine learning
algorithms namely the Back-Propagation Neural Network (BPNN) and the Online-
Sequential Extreme Learning Machine (ELM) presented in Sections 4.3.1 and 4.3.2
respectively are used as the learning machines in the framework in Figure 5.2.
6.2.1 Experimental Results of Compared Techniques
Comparing the developed framework (using SVM) with a traditional and a modern
ANN based approach, namely the BPNN (see Section 4.3.1) and the OS-ELM (see
Section 4.3.2), provides a better estimate of the memorization and generalization
capability of different learning machines.
The BPNN during training uses a different approach in the calculation of the
training error, as it minimizes the empirical error, whereas the SVM minimizes the
254
structural risk. Similar to the BPNN, the OS-ELM is a Single Layer Feed-forward
Neural Network (SLFN). Conventional ANN learning algorithms of SLFNs require
tuning of network parameters. However, the OS-ELM randomly generates the
input weights and the hidden neuron biases of the SLFN and uses them to calculate
the output weights without requiring further learning. The OS-ELM implemented
in this research is an online variant of the ELM algorithm, applicable for batch
learning (Liang et al., 2006).
Figure 6.6: Log-sigmoid transfer function
The network architecture of the BPNN implemented in this research consists of
1056 input neurons in the input layer, corresponding to the optimum subset of
1056 texture features (see Tables 5.6 and 5.7). The output layer of BPNN consists
of a single neuron, where an output of 0 indicates a benign sample and an output of
1 indicates a malignant sample. In the BPNN, the output of the neurons in the
hidden layers is calculated using the log-sigmoid activation function defined in
equation (4.46) and shown in Figure 6.6.
The number of samples selected for BPNN training and testing is indicated in Table
5.8. Three parameters need to be determined for the BPNN prior to obtaining a
trained model (classifier), which are as follows:
255
1. Number of hidden layers �a� 2. Number of hidden layer neurons �)� 3. Number of training iterations �!�
In order to statistically determine the optimum parameter values for the BPNN,
different combinations of parameter values of ), a and ! are iterated for the
following ranges: 1 ≤ a ≤ 20, 1 ≤ ) ≤ 1000 and 10 ≤ ! ≤ 2000. In order to
perform 10-fold CV, the 162 training samples (in Table 5.8) are split into CV
training and CV testing sets, such that 70 percent of the total samples from each
class are used for CV training (113 samples) and the remaining 30 percent samples
from each class are used for CV testing (49 samples). This procedure is repeated
for 100 trials using 10-fold CV, where on each trial CV training and CV testing
samples are selected randomly. The final structure of the BPNN after parameter
optimization results in a training accuracy of 93.58 percent (computed using
equation (5.21)), where the optimum BPNN parameters determined and used for
training are shown in Table 6.3. The BPNN classification results obtained after
testing the 70 samples (in Table 5.8) are shown in Table 6.4. The ROC curve
obtained from the BPNN classification results for testing with 70 samples is shown
in Figure 6.7 with an AUC of MN = 0.8235.
Table 6.3: Optimum parameters for the BPNN modeling
BPNN Parameters Optimum Value
Number of hidden layers a = 3
Number of hidden layer neurons
) = 409030�
where, ) is a matrix specifying the number of hidden neurons in each hidden layer of the BPNN.
Number of training iterations ! = 120
256
The OS-ELM is implemented in this research using the RBF activation function.
Using the RBF nodes, the centers and widths of the nodes are randomly generated
and fixed, based on this, the output weights are determined by the network.
The network architecture of the OS-ELM implemented in this research consists of
1056 input neurons in the input layer, corresponding to the optimum subset of
1056 texture features (see Section 5.4.3). The output layer of OS-ELM consists of a
single neuron, where an output of 0 indicates a benign sample and an output of 1
indicates a malignant sample. In the OS-ELM only one parameter needs to be
determined, which is the number of hidden layer neurons ), since the OS-ELM is a
SLFN. The method to search for the optimal number of the hidden layer neurons )
in the OS-ELM is suggested by (Huang et al., 2004), which indicates that the
number of hidden neurons, vary in the range from 20 to 200 as discussed in
Section 4.3.2.
Table 6.4: Comparison of the developed framework using different
machine learning techniques
Binary Classification
Performance Metrics SVM BPNN OS-ELM
TPs 35 30 33
FPs 1 6 3
FNs 1 5 4
TNs 33 29 30
Sensitivity 97.10% 85.71% 89.19%
Specificity 97.06% 82.86% 88.24%
Accuracy 97.14% 84.29% 90.00%
TPF 0.9710 0.8571 0.8919
FPF 0.0290 0.1429 0.1081
AUC �MN� 0.97574 0.8235 0.8971
257
The optimal value of ) is determined based on the classification performance of
the OS-ELM, which is the training accuracy (equation (5.21)). Since the number of
neurons in the input layer of the OS-ELM is large i.e., 1056, thus, for modeling
purposes, the range of ) is selected as 10 < ) ≤ 1000, where the size of ) is
incremented by a value of 10 on each iteration.
The final architecture of the OS-ELM after parameter optimization results in a
training accuracy (equation (5.21)) of 96.28 percent, where the optimal size of the
hidden layer neurons computed be ) = 160. The OS-ELM classification results
obtained after testing the 70 samples (in Table 5.8) are shown in Table 6.4. The
ROC curve obtained from the OS-ELM classification results for testing with 70
samples is shown in Figure 6.8 with an AUC of MN = 0.8971.
Figure 6.7: ROC curve of BPNN classifier for testing with 70 samples
(malignant is the + ! class and benign is the − ! class)
258
Figure 6.8: ROC curve of OS-ELM classifier for testing with 70 samples
(malignant is the + ! class and benign is the − ! class)
Parameter optimization for the BPNN and the OS-ELM in this research is
performed using 10-fold CV, which is similar to the case of the SVM. The reason for
using CV is that, since the number of training samples can be divided further into
subsets, CV ensures that the trained model (classification engine) does not overfit
the training data.
6.2.2 Discussion of Compared Models
Both SVMs and ANNs are considered as black-box modeling techniques. Although
both algorithms share the same structure, but the learning methods for both
algorithms are completely different. ANNs try to minimize the training error,
whereas SVMs reduce capacity using the SRM principle.
259
Comparison results of BPNN and OS-ELM in contrast to the SVM based model are
tabulated in Table 6.4, which are obtained for testing the 70 samples from the local
dataset (UMMC and MIAS). The experimental results in Table 6.4 show that the
SVM based approach outperforms the BPNN and the OS-ELM with respect to the
overall classification accuracy. This is because the optimum results for binary
classification are obtained by the SVM based model, where parameters: sensitivity,
specificity, TPF, FPF and MN are in optimum ranges.
Figure 6.9: ROC curves indicating the performance of the compared machine
learning techniques
To further investigate the accuracy of the compared machine learning models, the
ROC curves of all three models are computed using statistics from Table 6.4, as
shown in Figure 6.9. As observed from Figure 6.9, the SVM has the highest AUC
�MN = 0.97574�, followed by the OS-ELM �MN = 0.8971� and the BPNN
260
�MN = 0.8235�. The curve for the SVM follows the closest to the left-hand border
and then the top border of the ROC space this indicates that the SVM has better
classification results compared to the other techniques.
As observed from Table 6.4, the BPNN has the lowest performance in terms of the
classification accuracy out of all compared models. Since the BPNN used in this
research has a training accuracy of 93.58 percent, which is higher than the
generalization (testing) accuracy of 84.29 percent, this indicates that the BPNN has
a lower generalization as compared to the OS-ELM and the SVM. The main reason
for the low generalization of the BPNN is due to the cause of excessive training, i.e.
overfitting.
During BPNN training, the goal is to obtain a global optimum solution. However, in
a BPNN, to get the overall minimum answer of the error function, the network
extrema corrects itself slowly along the local improved way and eventually ends up
obtaining the local optimization answers only, which generally occurs due to
excessive training (overfitting). The reason for this is, since the BP algorithm is
based on the gradient descent approach, the network descends slowly with a low
learning speed and when a flat section (roof) appears for a long time the algorithm
ends the training at that instance, which results in locally optimized answers.
Another reason for the low generalization of the BPNN is due to noise in the digital
mammography data (features). The low generalization of the BPNN does not mean
that it is not a good tool for pattern classification, but given the reasons above, it is
not a considered a suitable tool to be evaluated with the mammography datasets
(in Table 5.1) acquired in this research.
261
In terms of classification performance, from Table 6.4 it is observed that the OS-
ELM ranks second after the SVM, whereby the BPNN ranks the last. The reason for
the better classification accuracy of the OS-ELM compared to the BPNN is that,
since the OS-ELM iteratively fine tunes the network’s input weights and biases
using finite samples of the training data, this yields a higher generalization for the
OS-ELM. The RBF transfer function applied in the OS-ELM technique randomly
initializes hidden neuron parameters such as the: input weight vectors, neuron
biases for additive hidden neurons, centers and impact factors for RBF hidden
neurons, and iteratively computes the output weight vectors.
During experimental testing of the OS-ELM technique it is observed that, if the
order of the training samples is switched or changed, the training accuracy of the
OS-ELM also changes significantly. In order to obtain an average estimate of the
memorization performance of the OS-ELM, the training accuracy of the OS-ELM is
computed using an average of 100 trials where on each trial training samples are
selected randomly. It is observed from the OS-ELM that, with the increase in the
number of input layer neurons, the OS-ELM achieves a better performance, while
remaining stable for a wide range of input neuron sizes.
There are a few reasons which constitute to the low performance of the OS-ELM
compared to the SVM. The first reason being that, the assignment of the initial
weights in the OS-ELM is arbitrary, which effects generalization performance of
network. As the proper selection of input weights and hidden bias values
contributes to the generalization capability of the trained model (classification
engine), the initialization of arbitrary weights decreases the generalization
performance of the OS-ELM.
262
The second reason being that, the value of � parameter in the RBF activation
function of the OS-ELM is set as a constant value of 1, as discussed by Huang et al.,
(2006b) and Liang et al. (2003). As the parameter � controls the width of the RBF
function in the OS-ELM, thus, it is suggested to be selected in within the range of 0
to 1. If the value of � is increased to � ≥ 1, the generalization performance for
unseen data will decrease.
More importantly, the value of � cannot be fixed to a constant value, since the
width of the Gaussian function depends upon the data samples to be classified and
also the amount of noise present in the data. Since there is no evidence or
literature on the ELM on how to tune the parameter � for the RBF activation
function, using the default parameter � = 1 as suggested by Huang et al., (2006b)
and Liang et al. (2003), the OS-ELM produces lower generalization performance
compared to the SVM. The OS-ELM does suffer from a few drawbacks, which are as
follows:
(a) For achieving good generalization results with the OS-ELM, the number of
hidden layer neurons �)� must be chosen larger than standard ANN
algorithms, (such as the BPNN). This is because the neuron weights and
biases are not learned from the training data.
(b) Multi-layer ANNs (such as the BPNN used in this research) if trained
properly, can possibly achieve similar and even better results comparable
to the OS-ELM, a SLFN.
(c) The solution provided by the ELM and the OS-ELM is not always so smooth,
and mostly shows some ripple.
263
The only notable advantage of the OS-ELM over the SVM is its faster training
process, with the increase in the chunk (data) size. It is known that using the RBF
(Gaussian) as the activation function, SVMs suffer from tedious parameter tuning.
However, the OS-ELM with a single parameter �)� to be tuned uses its arbitrary
assignment of initial random weights, which requires it to search for the optimal
size of hidden layer neurons �)�. This requires the OS-ELM to execute many times
in order to get an average estimate, which loses its edge over the SVM.
The experimental results presented in Section 6.2.1 indicate that using the SVM for
classification of malignant and benign abnormalities from digital mammography
data has shown to be very promising. In this research, SVMs have the a few notable
advantages as compared to ANNs, which are as follows:
• SVMs have non-linear dividing hypersurfaces that give them high
discrimination.
• They provide good generalization ability for unseen data classification.
• They determine the optimal network structure (such as the hidden layers
and hidden layer neurons) themself, without requiring to fine tune any
external parameters.
In contrast to the advantages of SVMs over ANNs, there are some drawbacks of
SVMs. However, these drawbacks are restricted due to practical aspects
concerning memory limitation and real-time training of SVMs. The drawbacks of
SVMs are as follows:
264
(a) The quadratic programming (QP) optimization problem arising in SVMs is
not easy to solve. Since the number of Lagrange multipliers is equal to the
number of training samples, the training process is relatively slow. Even
with the use of the Sequential Minimal Optimization (SMO), real-time
training is not possible for large datasets.
(b) The second drawback of SVMs is the requirement of storage capacity for the
trained model (classification engine). Support vectors (SVs) in the trained
model represent important features distinguishing the training samples
between the two classes (malignant and benign). When the optimization
problem has a low separability in the space used, the number of SVs
increases. SVs have to be stored in a model file. This puts limitations on the
implementation of SVM for devices with limited storage capacity.
Given all these aspects, the experimental results presented in Table 6.4 and Figure
6.9 shows that the SVM provides a better classification performance compared to
traditional and modern ANN based approaches. Thus, SVMs are considered as a
superior machine learning technique when the requirement is to solve
classification problems with noisy data.
6.3 Summary
This chapter presented the experimental results of the developed system in
Chapter 5. Section 6.1 presented and discussed the SVM training results relative to
the memorization and learning of the binary SVM classifier. Section 6.1 also
presented and discussed the SVM testing and validation results for unseen
samples. In order to perform a comparative research, Section 6.2 presented the
experimental results obtained after evaluating the developed framework using
265
different machine learning algorithms other than the SVM. The experimental
results of the compared machine learning models are discussed in the last part of
Chapter 6.
266
CHAPTER 7
CONCLUSION AND FUTURE WORK
7.0 Overview
This chapter concludes and summarizes the research contributions made. The
achievements and objectives of the research with respect to the experimental
results obtained are highlighted along with the key findings and significance of the
research. This chapter also discusses the impact and significance of the developed
system to radiologists and hospitals for mammography screening and
interpretation. Radiologists and clinicians will benefit from the developed system
as it will assist them in their diagnosis by acting as second readers.
7.1 Benefits of the Developed System
Digital mammography leads itself well to computerized detection of breast cancer,
where computer-aided methods based on image processing and machine learning
algorithms enable computers to identify suspicious areas of the breast that can be
mass lesions, MCCs or other signs of breast cancer.
In this research, an approach towards image processing and machine learning are
applied to develop a breast cancer detection system for classification between
malignant and benign abnormalities in digital mammograms, as shown in Figures
1.4 and 5.1. The modeling and development of the proposed system is presented in
Chapter 5. Firstly, the acquired digital mammography images in Table 5.1 are
preprocessed and segmented. Next, texture features are extracted from the
267
segmented mammogram images, where the optimal subset of features are selected
and classified using SVMs and ANNs. The experimental results presented in Section
6.2.2 (in Table 6.4) indicate that SVMs provide better classification performance
compared to traditional and modern ANNs.
Computerized breast cancer detection has provided a huge benefit in hospitals,
which are constantly looking for expert radiologists, and would have sensible
effects on medical and ethical grounds. Computer-aided methods have the
potential to increase the diagnostic accuracy by reducing the FPF, while increasing
the positive predictive values (PPVs) of mammographic abnormalities as discussed
in Chapter 3. The benefits obtained of using the developed system for breast
cancer detection are as follows:
1. This system will aid radiologist’s clinicians in the mammography screening
and interpretation process by acting as a second reader after the
radiologists.
2. This system will substantially reduce the number of false positives (FPs),
(see Section 3.3.6), which will eliminate the need of performing
unnecessary biopsies and save cost.
3. This system will reduce patient examination time by inspecting
mammograms and reporting the findings within a few seconds.
268
The weakest link in breast cancer detection has always been the radiologists, since
it is the radiologists who must find masses and MCCs. In cases where radiologists
have difficulties identifying between cancerous and non-cancerous abnormalities,
they can refer to this system for a second opinion as it is often difficult to
distinguish between malignant from benign abnormalities due to their similar
nature and visual features.
7.2 Contribution and Significance of Research
The detection of breast cancer in digital mammography applications using
machine learning approaches requires feature/heuristic computation (Woods and
Bowyer, 1996) from the Region of Interest (ROI), namely, the abnormal region. As
discussed in Section 1.3 previously, it is difficult to determine the size of the
neighbourhood (pixels) or the ROI that should be used to calculate the relevant
features from the abnormal regions (masses and/or MCCs). If the size of the ROI is
too large, small masses and/or MCCs may be missed, while if the size of the ROI is
too small, parts of large masses and/or MCCs may be missed. This poses a
challenging task in the computerized detection of breast cancer. Thus, the primary
contribution of this research as outlined in Section 1.3 is:
• To determine the most suitable ROI (neighbourhood) size of mass lesions
and MCCs for the purpose of feature computation (extraction).
The experimental results presented in Section 6.1.4.1.1 of this thesis contribute to
the problem of determining the optimum ROI size using the digital mammographic
data acquired in this research (in Table 5.1). Table 6.1 shows the experimental
results of evaluating different ROI sizes. It is observed from Table 6.1, that the ROI
269
with a size of 128 × 128 pixels achieves optimum results, with a classification
accuracy of 97.14 percent between malignant and benign samples. The
experimental results in Table 6.2 show that promising classification results can be
obtained by selecting the optimum ROI size for the purpose of feature
computation. The secondary contribution of this research outlined in Section 1.3
is:
• To demonstrate that advanced machine learning techniques, namely,
Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs)
can effectively solve pattern classification problems.
The secondary contribution of this research provides the basis for conducting a
comparative analysis between different machine learning technologies, such as
SVMs and ANNs. Since SVMs and ANNs are both learning machines, which share
the same structure but utilize different learning methods, two ANN approaches,
namely, the Back-propagation Neural Network (BPNN) (traditional approach) and
the Online-Sequential Extreme Learning Machine (OS-ELM) (modern approach)
are evaluated in this research for comparison with SVM.
The experimental results presented in Section 6.2.2 compares the three learning
machines: the BPNN, the OS-ELM and the SVM. As observed by the experimental
results in Table 6.4 and Figure 6.9, the SVM outperforms the BPNN and the OS-ELM
techniques in terms of classification performance. This indicates that the SVM has a
better generalization capability for the classification of malignant and benign
patterns as compared to traditional and modern ANN based techniques.
270
SVMs have a considerable advantage over ANNs, as they provide the use of soft
margins for the purpose of classification (see Section 4.2.4), thus allowing
improvement in the generalization performance of the developed system. In this
research, the observed advantages of SVMs over ANNs are as follows:
• SVMs have non-linear dividing hypersurfaces that give them high
discrimination capability, which is not the case with ANNs.
• SVMs provide good generalization ability for unseen data classification, as
they determine the optimal network structure themself, which is not the
case with ANNs.
• With the introduction of the SVM, the developed system is able to control
the balance between the sensitivity and the specificity, giving it more
flexibility.
The SVM classification engine developed has a good memorization capability. As
indicated from the experimental results in Section 6.1.3, the training accuracy of
the SVM (equation (5.21)) averaged over 100 trials is 97.60 percent, whereas the
10-fold CV accuracy is 87.83 percent. A training accuracy of greater than 95
percent indicates that the memorization and learning capability of the learning
machine is notably good, even with the presence of noisy data. The 10-fold CV
accuracy is taken as a measure to ensure the trained model (classification engine)
does not overfit the training data, where CV accuracy in between 80 to 95 percent
typically indicates good memorization capability with no overfitting. The two main
reasons contributing to the good training accuracy of the SVM classification engine
are:
271
1. Selection of the optimal subset of texture features for the learning machine
(SVM), as presented in Section 5.4.3.
2. Fine tuning of the SVM hyperplane parameters ��, �� using the -fold CV
approach, as presented in Section 5.5.3.1.
The experimental results obtained from testing the SVM classification engine with
unseen samples from the local (UMMC and MIAS) mammography datasets provide
a classification accuracy of 97.14 percent (in Table 6.4) for 70 testing samples.
Classification accuracy greater than 95 percent indicates promising results for
unseen data classification, for any learning machine.
It is known that the performance any learning machine can be problem dependent,
since the performance is based on a few factors such as: the experimental datasets
used, the optimum subset features selected for modeling and the method in which
the data samples are split between the training and testing sets. It is worth noticing
that SVMs have indicated lower classification performance compared to ANN
techniques, as reported by Osareh et al. (2002). Thus, the suitability of a learning
machine for a pattern classification task is data dependent. Since the local datasets
used in this research contain a noisy data, the SVM is found to be the most suitable
technique for classifying between malignant and benign patterns.
7.3 Achievement of Research Objectives
Digital mammography is a relatively new technique for the early detection of
breast cancer. It is based on accumulated density of tissues, i.e. to detect shadows.
This is the reason as to why mammography has been considered as an efficient
272
tool for the detection of masses and MCCs (Khuzi et al., 2009), (Verma & Zakos,
2001), (Jiang et al., 1999).
As discussed in Section 1.2, the goal of this research is to increase the diagnostic
accuracy of image processing and machine learning techniques for optimum
classification between malignant and benign abnormalities as well as to the
reproducibility of mammographic interpretation. In order to achieve this goal, the
research objectives outlined in Section 1.2 have been obtained, which are
discussed as follows:
1. The decision-logic system presented in Section 5.5.3.5 has shown to reduce
the number of false positives (FPs) (see Section 3.3.6). Moreover, from the
experimental results in Table 6.4 it is observed that SVMs are good are
reducing FPs. Thus, the techniques and algorithms implemented in this
research are capable of reducing the number of misclassified malignant
cancers (FPs), which complies with the third research objective of the
research.
2. The data acquired in this research is collected from different sources. The
acquired data is classified into two types, namely the local dataset and the
external dataset. The local dataset is a collection of digital mammography
images acquired from the University of Malaya Medical Centre (UMMC)
patient records. The external dataset is a well-known published image
database of 322 digital mammograms from the Mammographic Image
Analysis Society (MIAS) (in Section 5.2.1). Since two different datasets are
evaluated in this research, this complies with the fourth research objective.
273
3. To perform a comparative research, the experimental results obtained from
different learning machines are presented and compared in Chapter 6. The
proposed learning machine, the SVM, is compared to modern and
traditional ANN based techniques, namely the BPNN and the OS-ELM
respectively, as discussed in Section 6.2. The experimental results obtained
in Section 6.2.2 indicate that optimum performance for the classification
between benign and malignant patterns is obtained by the SVM technique.
This indicates the promising results of the SVM. This complies with the fifth
research objective.
7.4 Impact and Significance to Radiologists
The framework developed for the computerized breast cancer detection system in
Figure 5.1, can be implemented as an intelligent classification system to assist
radiologists in their diagnosis by acting as a second reader. The framework shown
in Figure 7.1 is termed as an Intelligent Classification System, which can be
implemented using the framework developed in this research. It is envisaged that
this system will aid radiologists in their interpretation of malignant and benign
abnormalities.
The intelligent classification system features a graphical user interface (GUI),
which allows radiologists to use the developed computerized breast cancer
detection system as a user-friendly software application. The intelligent
classification system can have up to three inputs (depending upon the nature of
diagnosis), and one output. Out of the three inputs, two inputs are compulsory for
diagnosis, which are as follows:
274
1. Digital mammogram image (image to diagnose/interpret).
2. Location of �O, P� co-ordinates of the center of the ROI (malignant and
benign abnormalities) to diagnose.
Figure 7.1: Intelligent classification system
The third input into the intelligent classification system can be taken as the
diameter of a circle enclosing the abnormality from the centre of the abnormality,
which is the location in �O, P� co-ordinates; the Ground Truth (GT) data (in Section
5.2.1). Since the value of the diameter optimally determines the size of the ROI
(neighborhood), the value of the diameter can be set as a variable parameter in the
Radiologist Suspicious digital
mammogram image
a. Location of (x,y)
co-ordinates of ROI
in image
b. Diameter of a
circle enclosing the
abnormality from
the centre of the
abnormality (x,y)
Image Processing
1. Image Preprocessing
2. Image Segmentation
Feature Normalization
Feature Extraction and
Selection
Region of Interest (ROI)
Selection
SVM Classification and
Diagnostic Results
SVM Binary
Classification
Result of Mammogram
Diagnosis
Trained SVM
Classifier
(Model)
Inputs into
Classification system
Standard process for
preparation of testing data
(Input)
(Output)
(Training performed
offline, i.e. once only)
Result: {Benign or Malignant}
275
GUI system, so it can be adjusted by radiologists based on their findings. Using the
ROI size as a variable parameter for the radiologists to specify will yield better
results with lesser false positives (FPs) (see Section 3.3.6). The output of the
intelligent classification system will indicate if the input mammogram is either a
benign or malignant abnormality.
Using the proposed intelligent classification system radiologists can incorporate
the output from the computer into their decision. Several recent studies
demonstrate that computer-aided detection improves radiologists’ ability in
differentiating malignant abnormalities from benign ones (Giger, 1999), (Jiang et
al., 1996), (Jiang et al, 1999), (Wu et al, 1993), (Huo et al., 2000), (D’Orsi et al.,
1992), (Baker et al, 1996), (Chan et al, 1999).
The framework developed in Figure 5.1 shows encouraging results as indicated in
Chapter 6, however, it also has a few limitations. These limitations can be
eliminated by implementing the future work suggested in the following section.
7.5 Future Expansion and Recommendations
Although the computerized breast cancer detection system can achieve a
classification accuracy of 97 percent as indicated in Chapter 6, however, this does
not guarantee it will obtain good and similar results on other mammography
datasets, especially those datasets which have not been tested in this research.
Thus, research can be further continued on to improve the performance of the
system and validating it by testing with larger digital mammography datasets such
as the Digital Database for Screening Mammography (DDSM) (Heath et al., 2001).
276
This researcher strongly believes that the computerized breast cancer detection
system developed in this research will contribute significant improvements in
mammographic interpretation of cancers. The following sections provide
suggestions and recommendations on future work that can be performed in order
to enhance the performance of the system and cater for untested mammography
datasets.
7.5.1 SVM Parameter Tuning using Genetic Algorithm
For any classification task, the performance of a learning machine will decrease if
the modelling (training) parameters are not selected properly. The modelling
parameters of a learning machine need to be fine-tuned (optimized) to obtain an
optimum balance between the generalization and memorization of the trained
model. Lagrangian parameter selection in the case of the SVM is complex in nature,
as it is difficult to solve by conventional optimization techniques (see Section
4.2.3.1).
A difficulty of using the SVM is the selection of parameter C and the kernel
parameter � (Gamma) in the RBF (Gaussian) kernel (in equation (4.44)). Even with
the use of -fold CV and the Grid-Search method (Hsu et al., 2003), an optimal
solution might not be achieved. The optimum values of the hyperparameters ��, �� in the SVM need to be found, so they can minimize the expectation of testing and
validation error, that needs to adapt to multiple parameters values at the same
time.
Genetic Algorithm (GA) with characteristics of high efficiency and global
optimization has been widely applied in many applications to solve optimization
277
problems (Anastasio et al., 1998). So, it is suggested that to solve the Dual
Lagrangian Optimization (DLO) problem in SVMs, the GA can be applied to
optimize the SVM hyperplane parameters ��, ��, This idea proposes a hybrid
combination of SVM-GA such that, the GA chromosomes will represent solutions as
the ��, �� parameters and the GA fitness function will evaluate the accuracy of the
solutions. The fittest (best) solutions obtained after the iterative GA process stops
will be the optimum ��, �� parameters for the SVM. The hybrid SVM-GA approach
suggested here will avoid the local optimum in finding the maximum Lagrangian.
7.5.2 Implementation of Multi-scale RBF Kernel
The RBF (Gaussian) kernel (in equation (4.44)) is one of the well-known Mercer’s
kernels (Mercer, 1909) for SVMs, which has been widely used in many
classification tasks, as the case of this research. The RBF kernel uses the Euclidean
distance between two points in the original space to find the correlation in the
augmented space. The points very close to each other are strongly correlated in the
augmented space, whereas the points far apart are uncorrelated in the augmented
space. There is only one parameter to adjust the width of the RBF kernel ± (sigma),
which is not powerful enough to cater for complex classification tasks.
The number of features used for machine learning modeling in this research is
1056 (see Tables 5.6 and 5.7), which is large. In order to achieve a better kernel for
SVMs using a large number of features, one possible way is to adjust the velocity of
decrement in each range of the Euclidean distance between the two points. The
multi-scale kernel obtained using this method will maintain the characteristics of a
RBF kernel. To implement this multi-scale kernel, a combination of RBF kernels at
different scales is suggested as the future work for this research. In order to
278
proceed, at first, a linear combination of the RBF must be satisfied to be Mercer’s
kernel. Theoretically, using a multi-scale RBF kernel, the performance and
classification accuracy of SVM will improve.
7.5.3 Evaluating Other Texture Approaches
This study uses GLCM texture descriptors for feature computation, as shown in
Table 5.7. The reason for using GLCMs in this research is due to the recent success
of GLCMs in digital mammography applications, as indicated by the literature
review in Table 3.3. However, this does not mean that amongst the texture based
techniques, GLCMs can only provide good feature computation results. Thus, other
texture based techniques can be evaluated in this research to perform a
comparative study. The following texture analysis techniques have gained recent
success in pattern classification problems and can be considered:
(a) Spatial Gray Level Dependence Method (SGLDM)
(b) Gray Level Run Length Matrix (GLRLM)
(c) Gray Level Difference Method (GLDM)
A comparative study evaluating other texture feature estimation approaches will
benefit this research, with a possibility in further improving the performance and
accuracy of the developed system.
7.6 Conclusion
In conclusion, this research has shown encouraging results and a performance that
matches human intelligence for classifying between malignant and benign
abnormalities in digital mammograms. The goal of this research has focused on
279
increasing the diagnostic accuracy of computer-aided detection methods used in
breast cancer detection.
The experimental results presented in Chapter 6 of this thesis, highlight the
significance and key contributions of this research which, the computerized breast
cancer detection system developed in this research has an average classification
accuracy of 95 percent for the SVM-based model. The system developed in this
research has a few notable advantages to radiologists. Firstly, this system will aid
clinical radiologists in the mammographic interpretation process by acting as a
second reader after the radiologists. Secondly, this system will reduce the number
of false positives (FPs), which will eliminate the need of performing unnecessary
biopsies and save costs. Lastly, this system will reduce patient examination time
by inspecting mammograms and reporting the findings within a few seconds.
As usually happens in an area of research, many approaches can be used and
developed, given the appropriate amount of time and effort. It is strongly
recommended that the future work suggested in Section 7.5 should be
investigated. The proper application and use of the developed system in this
research will be appreciated by radiologists in Malaysia. With the remarkable
pattern classification capability of SVMs, as shown in the experimental results in
Chapter 6, it is desired that more SVM-based applications should be developed for
the improvement and quality of health care systems.
280
REFERENCES
Adams, R. Bischof, L. 1994, Seeded Region Growing, IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 16, no. 6, pp. 641–647. Adler, D.D. and Helvie M.A. 1992, Mammographic biopsy recommendations, Curr.
Opin. Radiol., vol. 4, pp. 123–129. Alexander, F.E. Anderson, T.J. Brown, H.K. Forrest, A.P. Hepburn, W. Kirkpatrick,
A.E. Muir, B.B. Prescott, R.J. Smith, A. 1999. 14 years of follow-up from the Edinburgh randomised trial of breast-cancer screening, Lancet, vol. 353, pp. 1903-1908.
American Cancer Society, 2003a. Breast Cancer: Facts and Figures 2003-2004,
Atlanta, GA: American Cancer Society. American Cancer Society, 2003b. Cancer Prevention and Early Detection Facts and
Figures 2003-2004, Atlanta, GA: American Cancer Society. American Cancer Society, 2009. Breast Cancer: Facts and Figures 2009-2010,
Atlanta, GA: American Cancer Society. Anastasio, M.A. Yoshida, H. Nagel, R. Nishikawa, R.M. Doi, K. 1998, A genetic
algorithm-based method for optimizing the performance of a computer-aided diagnosis scheme for detection of clustered microcalcifications in mammograms, Medical Physics, vol. 25, no. 9, pp. 1613–1620.
Anttinen, I. Pamilo, M. Soiva, M. and Roiha, M. 1993, Double reading of
mammography screening films: one radiologist or two? Clinical Radiology, vol. 48, pp. 414-421.
Aylward, S.R.Hemminger, B.M. Pisano, E.D. Johnston, R.E. 1998, Mixture modeling
for digital mammogram display and analysis. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography, pp. 305–312. Kluwer, Dordrecht.
Baines, C.J. 1992, Breast self-examination. Cancer, vol. 69, pp. 1942–1946.
281
Bajger, M. Ma, F. and Bottema, M.J. 2005, Minimum Spanning Trees and Active
Contours for Identification of the Pectoral Muscle in Screening Mammograms, in Proc. of Digital Image Computing: Techniques and Applications, Dec. 2005, pp. 323–329.
Baker, J.A. Kornguth, P.J. Lo, J.Y. and Floyd Jr., C.E. 1996, Artificial neural network:
improving the quality of breast biopsy recommendations, Radiology, vol. 198, pp. 131–136.
Ballard, D.H. and Brown, C.M. 1982, Computer Vision. Englewood Cliffs, New Jersey:
Prentice-Hall. Bassett, L. W. Jackson, V. P. Jahan, R. Fu Y. S., and Gold, R. H. 1997, Diagnosis of
Diseases of the Breast, W. B. Saunders Company. Bazzani, A. Bollini, D. Campanini, R. Riccardi, A. Bevilacqua, A. Lanconelli, N.
Romani, D. 2001, Automatic detection of clustered microcalcifications using a combined method and an SVM classifier. In Yaffe, M.J. editor, Digital
Mammography 2000, in Proc. of the 5th International Workshop on Digital Mammography, Madison, Medical Physics Publishing, pp. 161–167.
Beam, C.A. Sullivan, D.C. 1994, What are the issues in the double reading of
mammograms?, Radiology, vol. 193, no. 2, pp. 582. Beam, C.A. Layde, P.M. Sullivan, D.C. 1996, Variability in the interpretation of
screening mammograms by u.s. radiologists, Arch Intern Med, vol. 156, pp. 209–213.
Behroozmand, R. Almasganj, F. 2005. Comparison of Neural Networks and Support
Vector Machines Applied to Optimized Features Extracted from Patients' Speech
Signal for Classification of Vocal Fold Inflammation”, in Proc. of the 5th IEEE Symposium on Signal Processing and Information Technology, pp. 844–849.
Beichel, R. Sonka, M. 2006. Computer vision approaches to medical image analysis.
Lecture Notes in Computer Science, vol. 4241, Springer, Berlin. Bellotti, R. De Carlo, F. Tangaro, S. Gargano, G. Maggipinto, G. Castellano, M.
Massafra, R. Cascio, D. Fauci, F. Magro, R. Raso, G. Lauria, A. Forni, G. Bagnasco, S. Cerello, P. Zanon, E. Cheran, S.C. Lopez Torres, E. Bottigli, U. Masala, G.L. Oliva, P. Retico, A. Fantacci, ME. Cataldo, R. De Mitri, I. De Nunzio, G. 2006, A completely automated CAD system for mass detection in a large mammographic database, Med. Phys., vol. 33, no. 8, August, pp. 3066-3075.
282
BI-RADS, Breast Imaging Reporting and Data System, American College of Radiology, 2010. [Online]. Website available at: http://www.acr.org/ SecondaryMainMenuCategories/quality_safety/BIRADSAtlas/BIRADSAtlasexcerptedtext/BIRADSMammographyFourthEdition.aspx
Bird, R.E. Wallace, T.W. Yankaskas, B.C. 1992, Analysis of cancers missed at
screening mammography, Radiology, vol. 184, pp. 613–617. Bishop, C. 1995, Neural Networks for Pattern Recognition: Oxford: Oxford
University Press. Bjurstam, N. Björneld, L. Duffy, S.W. Smith, T.C. Cahlin, E. Erikson,O. Lingaas, H.
Mattsson, J. Persson, S. Rudenstam, C.M. Säwe-Söderberg, J. 1997. The Gothenburg Breast Cancer Screening Trial: preliminary results on breast cancer mortality for women aged 39-49, J Natl Cancer Inst Monogr, vol. 22, pp. 53-55.
Blot, L. Zwiggelaar, R. 2000a. Extracting background texture in mammographic
images: a co-occurrence matrices based approach, in Proc. of the 5th International Workshop on Digital Mammography, Toronto, pp. 142–148.
Blot, L. Zwiggelaar, R. 2001. Background texture extraction for the classification of
mammographic parenchymal patterns, in Proc. of Medical Image Understanding and Analysis, pp. 145–148.
Blot, L. Zwiggelaar, R. Boggis, C.R.M. 2000b. Enhancement of abnormal structures
in mammographic images, in Proc. of Medical Image Understanding and
Analysis, pp. 125–128. Blot, L. Davis, A. Holubinka, M. Martì, R. Zwiggelaar, R. 2002. Automated quality
assurance applied to mammographic imaging, EURASIP Journal Applied Signal
Processing, vol. 2002, no. 1, pp. 736–745. Boser, B. Guyon, I. Vapnik, V. 1992. A training algorithm for optimal margin
classifiers, in Proc. of the 5th Annual ACM Workshop on Computation Learning Theory, pp. 144–152, ACM Press, Pittsburgh.
Bottema, M.J. Slavotinek, J.P. 1998, Detection of subtle microcalcifications in digital
mammograms. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography 1998, pp. 209–212. Kluwer, Dordrecht.
283
Bottema, M.J. Slavotinek, J.P. 2001, Detection of microcalcifications associated with cancer. In Yaffe, M.J. editor, Digital Mammography 2000, in Proc. of the 5th International Workshop on Digital Mammography, Madison, Medical Physics Publishing, pp. 149–153.
Bovis, K. Singh, S 2000. Detection of masses in mammograms using texture features,
in Proc. of the 15th International Conference on Pattern Recognition, vol. 2, 2267–2270.
Bovis, K. Singh, S. 2002. Classification of mammographic breast density using a
combined classifier paradigm, in Proc. of the 4th International Workshop on Digital Mammography, pp. 177–180.
Boyle, P. and Levin, B. 2008, World Cancer Report 2008, International Agency for
Research on Cancer, Lyon, France. Bozek, J. Delac, K. and Grgic, M. 2008, Computer-Aided Detection and Diagnosis of
Breast Abnormalities in Digital Mammography, Proceedings of the 50th International Symposium ELMAR-2008, Zadar, Croatia, pp. 45-52.
Brieman, L. 2001. Random Forests, Machine Learning, vol. 5, no. 1, pp. 5–32. Brown, S. Li, R. Brandt, L. Wilson, L. Kossoff, G. Kossoff, M. 1998, Development of a
multi-feature CAD system for mammography. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography, pp. 189–196. Kluwer, Dordrecht.
Brown, J. Bryan, S. Warren, R. 1996, Mammography screening: an incremental cost
effectiveness analysis of double versus single reading of mammograms, BMJ, vol. 312, no. 7034, pp. 809–812.
Bruynooghe, M. 2001, High resolution granulometric analysis for early detection of
small microcalcification clusters in X-ray mammograms. In Yaffe, M.J. editor, Digital Mammography 2000, in Proc. of the 5th International Workshop on Digital Mammography, Madison, Medical Physics Publishing, pp. 154–160.
Brzakovic, D. Luo, X.M. Brzakovic, P. 1990, An approach to automated detection of
tumors in mammograms, IEEE Transactions on Medical Imaging, vol. 9, no. 3, pp. 233–241.
284
Burges, C.J.C. 1988. A tutorial on support vector machines for pattern recoginition. Data Mining and Knowledge Discovery, vol. 2, pp. 121–167.
Burrel, H.C. Sibbering, D.M. Wilson, A.R.M. Pinder, S.E. Evans, A.J. Yeoman, L.J.
Elston, C.W. Ellis, I.O. Blamey, R.W. Robertson, J.F.R. 1996, Screening interval breast cancers: mammographic features and prognostic factors, Radiology, vol. 199, pp. 811–817.
Buseman, S. Mouchawar, J. Calonge, N. Byers, T. 2003, Mammgraphy screening
matters for young women with breast carcinoma, Cancer, vol. 97, no. 1, pp. 352-358.
Byng, J.W. Critten, J.P. Yaffe, M.J. 1997, Thickness-equalization processing for
mammographic images. Radiology, vol. 203, pp. 564–568. Byvatov, E. Fechner, U. Sadowski, J. Schneider, G. 2003. Comparison of Support
Vector Machine and Artificial Neural Network Systems for Drug/Nondrug Classification, Journal of Chemical Information and Computer Science, vol. 43, no. 6, pp. 1882–1889.
Cascio, D. Fauci, F. Magro, R. Raso, G Bellotti, R. De Carlo, F. Tangaro, S. De Nunzio,
G. Quarta, M. Forni, G. Lauria, A. Fantacci, M.E. Retico, A. Masala, G.L. Oliva, P. Bagnasco, S. Cheran, S.C. and Torres, E.L. 2008, ‘Mammogram Segmentation by Contour Searching and Mass Lesions Classification With Neural Network’, IEEE Transactions on Nuclear Science, vol. 53, no. 5, pp. 2827-2833.
Cardenosa, G. 1996, Mammography: An overview. In Doi, K. Giger, M.L. Nishikawa,
R.M. Schmidt, R.A. editors, Digital Mammography, pp. 3–10. Elsevier, Amsterdam.
Cernadas, E. Gomez, L. Rodriguez, P.G. Casas, A. Carrion, R.G. Vidal, J.J. 1996, Design
of unsharp masking filters in the frequency domain: Parametrization for breast radiographs. In Doi, K. Giger, M.L. Nishikawa, R.M. Schmidt, R.A. editors, Digital Mammography, pp. 463–466. Elsevier, Amsterdam.
Cernadas, E. Zwiggelaar, R. Veldkamp, W. Parr, T. Astley, S. Taylor, C. Boggis, C.
1998, Detection of mammographic microcalcifications using a statistical model. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography 1998, pp. 205–208. Kluwer, Dordrecht.
285
Chan, H.P. Doi, K. Galhotra, S. Vyborny, C.J. MacMahon, H. Jokich, P.M. 1987, Image feature analysis and computer-aided diagnosis in digital radiography. I. Automated detection of microcalcifications in mammography, Med Phys, vol. 14, no. 4 pp. 538–48.
Chan, H.P. Doi, K. Vyborny, C.J.1990. Improvements in radiologists' detection of
clustered microcalcifications on mammograms: the potential of computeraided diagnosis, Investigative Radiology, vol. 25, pp. 1102–1110.
Chan, H. P. Sahiner, B. Lam, K.L. Petrick, N. Helvie, M.A. Goodsitt, M.M. and Adler,
D.D. 1998, ‘Computerized analysis of mammographic microcalcifications in morphological and texture feature spaces’, Medical Physics, vol. 25, no. 10, pp. 2007–2019.
Chan, H.P. Sahiner, B. Helvie, M.A. Petrick, N. Roubidoux, M.A. Wilson, T.E. Adler,
D.D. Paramagul, C. Newman, J.S. Sanjay-Gopal, S. 1999, Improvement of radiologists’ characterization of mammographic masses by using computer-aided diagnosis: an ROC study, Radiology, vol. 212, pp. 817–827, 1999.
Chang, C.-C. and Lin, C.-J. 2010, LIBSVM: A library for support vector machines.
[Online]. Available at: http://www.csie.ntu.edu.tw/~cjlin/libsvm Chen, S.-Y. Lin, W.-C. Chen, C.-T. 1991, Split-and-merge image segmentation based
on localized feature analysis and statistical tests, CVGIP: Graphic. Models
Image Processing. vol. 53, no. 5, pp. 457–475. Cheng, H-.D. Lui, Y.M. Freimanis, R.I. 1998, A novel approach to microcalcification
detection using fuzzy logic technique, IEEE Trans on Medical Imaging, vol. 17, no. 3, pp. 442–450.
Chen, Y.W. and Lin, C.J. 2006, Combining SVMs with various feature selection
strategies. In Guyon, I. Gunn, S. Nikravesh, M. Zadeh, L., editors, Feature
Extraction, Foundations and Applications, Springer, New York. Cheng, H.D. Shi, X.J. Min, R. Hu, L.M. Cai, X.P. Du, H.N. 2006, Approaches for
automated detection and classification of masses in mammograms, Pattern
Recognition, vol. 39, no. 4, pp. 646-668. Cheevasuvit, F. Maitre, H. Vidal-Madjar, D. 1986, A robust method for picture
segmentation based on split-and-merge procedure, Comput. Vis. Graph. Image
Process.. vol. 34, pp. 268–281.
286
Chitre, Y. Dhawan, A.P. Moskowitz, M. 1994, Classification of mammographic microcalcifications using image structure and cluster features. In Gale, A.G. Astley, S.M. Dance, D.R. Cairns, A.Y. editors, Digital Mammography, pp. 31–40. Elsevier, Amsterdam.
Clausi, D.A. 2002. An analysis of co-occurrence texture statistics as a function of
grey level quantization, Canadian Journal of Remote Sensing, vol. 28, no. 1, pp. 45–62.
Cortes, C. Vapnik, V. 1995. Support vector networks. Machine Learning, vol. 20, pp.
273–297. Costaridou, L. Sakellaropoulos, P.N. Kristalli, M.A. Skiadopoulos, S.G. Karahaliou,
A.N. Boniatis, I.S. Panayiotakis, G.S. 2005. Multiresolution feature analysis for
differentiation of breast masses from normal tissue, in Proc. of the 1st International Conference on Experiments/Process/System/Modelling and Optimization, Greece, Athens.
Chu, K.C., Smart, C.R. Tarone, R.E. 1988. Analysis of breast cancer mortality and
stage distribution by age for the Health Insurance Plan clinical trial, Journal of
the National Cancer Institute, vol. 14, pp. 1125-1132. D’Orsi, C.J. Getty, D.J. Swets, J.A. Pickett, R.M. Seltzer, S.E. and McNeil, B.J. 1992,
Reading and decision aids for improved accuracy and standardization of mammographic diagnosis, Radiology, vol. 184, pp. 619–622.
Dash M. and Liu, H. 1997, Feature selection for classification, Intelligent Data
Analysis, vol. 1, no. 3, pp. 131–156. Day, W.H.E. Edelsbrunner, H. 1984. Efficient algorithms for agglomerative
hierarchical clustering methods, Journal of Classification, vol. 1, pp. 7–24. Davies, D.H. Dance, D.R. 1990. Automatic computer detection of clustered
calcifications in digital mammograms, Phys. Med. Biol., vol. 35, pp. 1111–1118. De Koning, H.J. Fracheboud, J. Boer, R. Verbeek, A.L. Collette, H.J. Hendriks, J.H.C.L
van Ineveld, B.M. de Bruyn A.E., and van der Maas P.J. 1995, Nation-wide breast cancer screening in Netherlands: support for breast cancer mortaility reduction. National evaluation team for breast cancer screening, International
Journal of Cancer, vol. 60, no. 6, pp. 777-780.
287
Dean, P.B. 1996, Overview of breast cancer screening. In Doi, K. Giger, M.L. Nishikawa, R.M. Schmidt, R.A. editors, Digital Mammography, pp. 19–26. Elsevier, Amsterdam.
Degroeve, S. Tanghe, K. Baets, B.D. Leman, M. Martens, J.-P. 2005. A Simulated
Annealing Optimization of Audio Features for Drum Classification, in Proc. of the 6th International Conference on Music Information Retrieval, London, pp. 482–487.
Dehghan, F. Abrishami Moghaddam, H. Giti, M. 2008, "Automatic Detection of
Clustered Microcalcifications in Digital Mammograms Study on Applying Adaboost with SVM-based Component Classifiers", Proceedings of the 30th Annual International IEEE EMBS Conference, Vancouver, Canada, pp. 4789-4792.
Dengler, J. Behrens, S. Desaga, J.F. 1993, Segmentation of microcalcifications in
mammograms, IEEE Transactions on Medical Imaging, vol. 12, no. 4, pp. 634–642.
Dept. of Biomedical Imaging, Faculty of Medicine, University of Malaya (UM), 2010.
[Online]. Available at: http://radiology.um.edu.my/ Dept of Biomedical Imaging, University Malaya Medical Centre (UMMC), Kuala
Lumpur, 2010. [Online]. More information available at: http://www.ummc.edu.my/index.php?option=com_content&view=article&id =51:medical-imaging-services&catid=39clinical service s&Itemid=55
Diahi, J.G. Giron, A. Brahmi, D. Frouge, C. Fertil, B. 1998, Evaluation of a neural
network classifier for detection of microcalcifications and opacities in digital mammograms. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography 1998, pp. 151–156. Kluwer, Dordrecht.
Diyana, W.M. Zulaikha, K. and Besar, R. 2002, An Intelligent CAD System for Breast
Cancer Detection in Digital Mammograms, International Conference on Artificial Intelligence in Engineering & Technology, Kota Kinabalu Sabah, pp. 476-479.
Diyana, W.M. Besar, R. 2003a, Automated Methods in Clustered Microcalcifications
Detection Module of a CAD System, World Scientific Journals: Journal of
Mechanics in Medicine and Biology, vol. 3, no. 3, pp. 30-33.
288
Diyana, W.M. Larcher, J. and Besar, R. 2003b, A Comparison of Clustered Microcalcifications Automated Detection Methods in Digital Mammograms. IEEE International Conference on Accoustic, Speech and Signal Processing, pp. 385-388.
Doi, K. Giger, M.L. Nishikawa, R.M. Hoffmann, K.R. MacMahon, H. Schmidt, R. Chua,
K.G. 1993, Digital radiography: a useful clinical tool for computer-aided diagnosis by quantitative analysis of radiographic images, Acta Radiologica, vol. 34, pp. 426–439.
Domínguez, A.R. Nandi, A.K. 2008. Detection of masses in mammograms via
statistically based enhancement, multilevel-thresholding segmentation, and region selection, Computerized Medical Imaging and Graphics, vol. 32, no. 4, pp. 304–315.
Domínguez, A.R. Nandi, A.K. 2009a. Toward breast cancer diagnosis based on
automated segmentation of masses in mammograms, Pattern Recognition, vol. 24, no. 6, pp. 1138–1148.
Domínguez, A.R. Nandi, A.K. 2009b. Development of tolerant features for
characterization of masses in mammograms, Computers in Biology and
Medicine, vol. 39, no. 8, pp. 678–688. Dror, G. Sorek, R. Shamir, S. 2005. Accurate Identification of Alternatively Spliced
Exons using Support Vector Machine, Bioinformatics, vol. 21, no. 7, pp. 897–901.
Dua, S. Singh, H. Thompson, H.W 2009. Associative classification of mammograms
using weighted rules, Expert Systems with Applications, vol. 36, no. 5, pp. 9250–9259.
Duda, R.O. Hart, R.E. Stork, D.G. 2001, Pattern Classification. John Wiley & Sons. Dumais, S. 1998. Using SVMs for Text Categorization, IEEE Intelligent Systems, vol.
13, pp. 21–23. El-Naqa, I. Yang, Y. Wernick, M.N. Galatsanos, N.P. Nishikawa, R.M. 2002, A Support
Vector Machine Approach for Detection of Microcalcifications, IEEE
Transactions on Medical Iamging, vol. 21, no. 12, pp. 1552-1563.
289
Elmore, J.G. Wells, C.K. Lee, C.H. Howard, D.H. Feinstein. A.R. 1994, Variability in radiologists’ interpretations of mammograms, N Engl J Med, vol. 331, no. 22, pp. 1493–1499.
Esteve, J. Kricker, A. Ferlay, J. and Parkin, D. 1993, Facts and figures of cancer in the
European Community, Technical report, International Agency for Research on Cancer, Lyon, France.
Evans, W.P. 1995, “Breast Masses Appropriate Evaluation”, The Radiologic Clinics of
North America, Breast Imaging, vol. 33, no. 6, pp. 1085-1108. Feature selection tool for LIBSVM written in python language, 2010. [Online].
Available at: http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/ Feig, S.A. Yaffe, M.J. 1995, Digital Mammography, Computer-Aided Diagnosis and
Telemammography, The Radiologic Clinics of North America, Breast Imaging, vol. 33, no. 6, pp. 1205-1230.
Ferlay, J. Autier, P. Boniol, M. Heanue, M. Colombet, M., Boyle, P. 2007, Estimates of
the cancer incidence and mortality in Europe in 2006, Annals of Oncology, vol. 18, no. 5, pp 581-592.
Ferrari, R.J. Rangayyan, R.M. Desautels, J.E.L. Borges, R.A. and Frere, A.F. 2004,
Automatic identification of the pectoral muscle in mammograms, IEEE
Transactions on Medical Imaging, vol. 23, no. 2, pp. 232–245. Ferrari, R.J. Rangayyan, R.M. Desautels, J.E.L. Borges, R.A. and Frere, A.F. 2001,
Analysis of asymmetry in mammograms via directional filtering with Gabor wavelets, IEEE Transactions on Medical Imaging, vol. 20, no. 9, pp. 953–964.
Fisher, R.A. 1936, The use of multiple measurements in taxonomic problems, Ann.
Eugenics, vol. 7, pp. 178-188. Fogel, D.B. Wasson, E.C. Boughton, E.M. Porto, V.W. Angeline, P.J. 1998, Linear and
neural models for classifying breast masses, IEEE Transactions on Medical
Imaging, vol. 17, no. 3, pp. 485-488. Frieß, T.-T. N. Cristianini, N. Campbell, C. 1998. The Kernel-Adatron Algorithm: A
Fast and Simple Learning Procedure for Support Vector Machines, in Proc. of the 15th International Conference on Machine Learning, San Francisco, California, pp. 188–196.
290
Friedman, P.J. 1999, The past and future of radiological error. In Krupinski, E.A. editor, Medical imaging 1999: Image perception and performance, vol. 3663, pp. 2–7.
Friedrich, M. Sickles, E.A. (eds.) 2000. Radiological diagnosis of breast diseases.
Springer. Frisell, J. Lidbrink, E. Hellström, L. Rutqvist, L.E. 1997. Followup after 11 years–
update of mortality results in the Stockholm mammographic screeningtrial, Breast Cancer Res Treat, vol. 45, no. 3, pp. 263-270.
Fukunaga, K. 1990, Introduction to Statistical Pattern Recognition. Academic Press. Fukuoka, D. Kasai, S. Fujita, H. Hara, T. Kato, M. Endo, T. Yoshimura, H. 1998,
Automated detection of clustered microcalcifications on digitized mammograms. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography 1998, pp. 197–200. Kluwer, Dordrecht.
Furey, T.S. Christiani, N. Duffy, N. Bednarski, D.W. Schummer, M. Hauessler, D.
2000. Support vector machine classification and validation of cancer tissue samples using microarray expression data, Bioinformatics, vol. 16, pp. 906–914.
Glatt, L.A. Longbotham, H.G. Arnow, T.L. Shelton, D. Ravdin. P. 1992, Application of
weighted-majority minimum-range filters in the detection and sizing of tumors in mammograms. In Loew, M.H. editors, Medical Imaging VI: Image
Processing, pp. 477–488. Giger, L. Yin, F. Doi, K. 1990. Investigation of methods for the computerised
detection and analysis of mammographic masses, SPIE Medical Imaging and
Image Processing IV, vol. 1233, pp. 183–184. Giger, M.L. Vyborny, C.J. Schmidt , R.A. 1994, Computerized characterization of
mammographic masses: analysis of spiculation, Cancer Letters, vol. 77, pp. 201–211.
Giger, M. MacMahon, H. 1996, Image processing and computer-aided diagnosis,
Radiologic Clinics of North America, vol. 24, no. 3, pp. 565-596.
291
Giger, M.L. 1999, Overview of computer-aided diagnosis in breast imaging. In: Computer Aided Diagnosis in Medical Imaging, (Doi K, MacMahon H, Giger ML, Hoffmann KR, eds). (Elsevier, Amsterdam), pp. 167-176.
Goergen, S.K. Evans, J.E. Cohen, G.P.B. Macmillan, J.H. 1997. Characteristics of
breast carcinomas missed by screening radiologists, Radiology, vol. 204, pp. 131-135.
Golub, T.R. Slonim, D.K. Tamayo, P. Huard, C. Gaansenbeek, M. Mesirov, J.P. Coller,
H. Loh, M.L. Downing, J.R. Caliguiri, M.A. Bloomfield, C.D. Lander, E.S. 1999. Molecular classification of cancer: Class discovery and class prediction by gene expression monitoring, Science, vol. 286, pp. 531–557.
Gonzalez, R.C. and Woods, R.E. 2002, Digital Image Processing. Prentice-Hall, Inc.,
2nd edition, ISBN 0-201-18075-8. International Edition. Gotzsche, P.C. Olsen, O. 2000. Is screening for breast cancer with mammography
justifiable?, Lancet, vol. 355, pp. 129-134. Gorgel, P. Sertbas, A. Kilic, N. Ucan, O.N. Osman, O. 2009, Mammographic Mass
Classification using Wavelet Based Support Vector Machine, Istanbul
University - Journal of Electrical and Electronics Engineering, vol. 9, no. 1, pp. 867-875.
Green, D.M. and Swets, J.A. 1966, Signal detection theory. Wiley, New York. Groshong, B.R. Kegelmeyer, W.P. 1996, Evaluation of a hough transform method
for circumscribed lesion detection. In Doi, K. Giger, M.L. Nishikawa, R.M. Schmidt, R.A. editors, Digital Mammography, pp. 361–366. Elsevier, Amsterdam.
Guillemet, H. Benali, H. Kahn, E. DiPaola, R. 1996, Detection and characterization of
microcalcifications in digital mammography. In Doi, K. Giger, M.L. Nishikawa, R.M Schmidt, R.A. editors, Digital Mammography 1996, pp 225–230. Elsevier, Amsterdam.
Guliato, D. Rangayyan, R.M. de Carvalho, J.D. Santiago, S. A. 2006, "Spiculation-
Preserving Polygonal Modeling of Contours of Breast Tumors", in Proceedings of the 28th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBS), New York, August-September 2006, pp. 2791-2794.
292
Gulsrud, T.O. Loland, E. 1996. Multichannel filtering for texture extraction in digital mammograms, in Proc. of the 18th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society, Amsterdam, The Netherlands.
Gupta, R. Undrill, P.E. 1995. The use of texture analysis to delineate suspicious
masses in mammography, IOPscience, vol. 40, pp. 835–855. Guyon, I. Boser, B. Vapnik, V. 1993. Automatic capacity tuning of very large VC-
dimension classifiers, Advances in Neural Information Processing Systems, vol. 5 Morgan Kaufmann, San Mateo, CA.
Guyon, I. Eliseeff A. E. 2003, An introduction to variable and feature selection.
Journal of Machine Learning Research, vol. 3 pp. 1157–1182. Gürcan, M.N. Yardimci, Y. Cetín, A.E. 1998, Microcalcifications detection using
adaptive filtering and Gaussianitiy tests. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography 1998, pp. 157–164. Kluwer, Dordrecht.
Hadjiiski, L. Sahiner, B. Chan, H.P. Petrick, N. Helvie, M. 1999, Classification of
Malignant and Benign Masses Based on Hybrid ART2LDA Approach, IEEE
Transactions on Medical Imaging, vol. 18, no. 12, pp. 1178-1187 Hagihara, Y. Kobatake, H. Nawano, S. Takeo, H. 2001, Accurate detection of
microcalcifications on mammograms by improvement of morphological processing. In Yaffe, M.J. editor, Digital Mammography 2000, in Proc. of the 5th International Workshop on Digital Mammography, Madison, Medical Physics Publishing, pp. 193–197.
Hara, T. Yamada, A. Fujita, H. Iwase, T. Endo, T. 2001, Automated classification
method of mammographic microcalcifications by using artificial neural network and ACR BI-RADStm Criteria for Microcalcification Distribution. In Yaffe, M.J. editor, Digital Mammography 2000, in Proc. of the 5th International Workshop on Digital Mammography, Madison, Medical Physics Publishing, pp. 198–204.
Haralick, R.M. Shanmugam, K. and Dinstein, I., 1973, Textural Features for Image
Classification, IEEE Transactions on Systems, Man and Cybernetics―Part C, vol. 3, no. 6, pp. 610–621.
293
Haralick, R.M. 1979. Statistical and structural approaches to texture, Proceedings of
the IEEE, vol. 67, pp. 786–804. Hartswood, M. Procter, R. Williams, L.J. 1998, Prompting in practice: How can we
ensure radiologists make best use of computer-aided detection systems in screening mammography. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. Van Erning, L.J.T.O. editors, Digital Mammography, pp. 363–370. Kluwer, Dordrecht.
Harris, R. Kinsinger, L.S. 2002, Routinely Teaching Breast Self-Examination is Dead.
What Does This Mean?, Journal of the National Cancer Institute, vol. 94, no. 19, pp. 1420-1421.
Harris, J.R. Henderson, C.I. Hellman, S. Kinne, D.W. 1996, Breast Diseases, 2nd
edition, JB Lippincott Company, Philadelphia, PA. Harvey, J.E. Fajardo, L.L. Inis, C.A. 1993, Previous mammograms in patients with
impalpable breast carcinoma: retrospective vs blinded interpretation, AJR, vol. 161, pp. 1167–1172.
Haykin, S. 1999. Neural Networks: A Comprehensive Foundation. Prentice Hall. Hassanien, A. Slezak, D. 2006. Rough neural intelligent approach for image
classification: A case of patients with suspected breast cancer, International
Journal of Hybrid Intelligent Systems, vol. 3, pp. 205–218. Hassanien, A. 2007. Fuzzy rough sets hybrid scheme for breast cancer detection,
Image and Vision Computing, vol. 25, no. 2, pp. 172–183. Heath, M. Bowyer, K. Kopans, D. Moore, R. Kegelmeyer, W.P. 2001, The Digital
Database for Screening Mammography, in Proceedings of the Fifth International Workshop on Digital Mammography, Yaffe, M.J. ed., Medical
Physics Publishing, 2001, pp. 212–218. Hendee, W.R. Beam, C. and Hendrick, E. 1999, Proposition: all mammograms
shouyld be double-read, Medical Physics, vol. 26, pp. 115-118. Highnam, R.P. Brady, J.M. Shepstone, B.J. 1996, Removing the anti-scatter grid in
mammography. In Doi, K. Giger, M.L. Nishikawa, R.M Schmidt, R.A. editors, Digital Mammography 1996, pp 459–462. Elsevier, Amsterdam.
294
Highnam, R. and Brady M. 1999, Mammographic Image Analysis, Kluwer Academic Publishers, Dordrecht, The Netherlands.
Homer, M.J. 1997. Mammographic interpretation. The McGraw-Hill Companies, Inc. Horowitz, S. L. Pavlidis, T. 1974, Picture segmentation by a directed split-and-merge
procedure, in Proc. of the 2nd International Joint Conference on Pattern Recognition, pp. 424–433.
Howard, D. Roberts, S.C. Ryan, C. Brezulianu, A. 2008. Textural classification of
mammographic parenchymal patterns with the sonnet selforganizing neural network, Journal of Biomedicine and Biotechnology, vol. 2008, Article ID 526343, 11 pages.
Hsu, C.-N. Huang, H.-J. Dietrich, S. 2002, The ANNIGMA- wrapper approach to fast
feature selection for neural nets. IEEE Trans Systems, Man and Cybernetics,
vol. 32, no. 2, pp. 207–212. Hsu, C.-W. Chang, C.-C. Lin, C.-J. 2003. A Practical Guide to Support Vector
Classification. Technical Report, Department of Computer Science and Information Engineering, National Taiwan University, Taipei, 2003.
Huang, G.-B. Babri, H.A. 1998, Upper Bounds on the Number of Hidden Neurons in
Feedforward Networks with Arbitrary Bounded Nonlinear Activation Functions, IEEE Transactions on Neural Networks, vol. 9, no. 1, pp. 224–229.
Huang, G.-B. 2003, Learning Capability and Storage Capacity of Two-Hidden-Layer
Feedforward Networks, IEEE Transactions on Neural Networks, vol. 14, no. 2, pp. 274–281.
Huang, S.F. Chang, R.F. Chen, D.R. Moon, W.K. 2004, Characterization of spiculation
on ultrasound lesions, IEEE Transactions on Medical Imaging, vol. 23, no. 1, pp. 111-121.
Huang, G.-B. Zhu, Q.-Y. Siew, C.-K. 2004, Extreme Learning Machine: A New Learning
Scheme of Feedforward Neural Networks, in Proc. of the International Joint Conference on Neural Networks, Budapest, Hungary, vol. 2, pp. 985–990.
Huang, G.-B. and Chee-Kheong, S. 2004, Extreme Learning Machine: RBF Network
Case, in Proc. of the 8th International Conference on Control, Automation, Robotics and Vision, Kunming, China, vol. 2, pp. 1029–1036.
295
Huang, G.-B. Zhu, Q.-Y. Siew, C.-K. 2006a, Extreme Learning Machine: Theory and
Applications, Neurocomputing, vol. 70, no. 1-3, pp. 489–501. Huang, G.-B. Chen, L. Siew, C.-K. 2006b, Universal Approximation using
Incremental Constructive Feedforward Networks with Random Hidden Nodes, IEEE Transactions on Neural Networks, vol. 17, no. 4, pp. 879–892.
Huang, G.-B. Zhu, Q.-Y. Mao, K. Z. Siew, C.-K. Saratchandran, P. Sundararajan, N.
2006c, Can Threshold Networks be Trained Directly?, IEEE Transactions on
Circuits and Systems - II: Express Briefs, vol. 53, no. 3, pp. 187–191. Huo, Z. Giger, M.L. Vyborny, C.J. Bick, U. Lu, P. Wolverton, D.E. Schmidt, R.A. 1995,
Analysis of spiculation in the computerized classification of mammographic masses, Med Phys, vol. 22, pp. 1569–1579.
Huo, Z. Giger, M.L. Vyborny, C.J. and Metz, C.E. 2000, Effectiveness of CAD in the
diagnosis of breast cancer: an observer study on an independent database of mammograms, Radiology, vol. 7, pp. 1077–1084.
Hutt, I. 1996, The computer-aided detection of abnormalities in digital
mammograms. PhD thesis, Faculty of Medicine, Department of Medical Biophysics, University of Manchester.
IARC Handbooks of Cancer Prevention 2002, Beast Cancer Screening, vol. 7, Lyon:
IARC Press 2002. Ibrahim, N. Fujita, H. Hara, T. Endo, T. 1997. Automated detection of clustered
microcalcifications on mammograms: CAD system application to MIAS database, Physics in Medicine and Biology, vol. 42, no. 12, pp. 2577–2589.
Image Processing Toolbox (R2009b), MATLAB (Matrix Laboratory). Mathworks,
2009. [Online]. Available at: http://www.mathworks.com/help/toolbox /images/
Jackson, V.P. 2002. Screening mammography: controversies and headlines.
Radiology, vol. 225, no. 2, pp. 323-326.
296
Jiang, Y. Nishikawa R.M. Wolverton, E.E. Metz, C.E. Giger, M.L. Schmidt, R.A. and Vyborny, C.J. 1996, Malignant and benign clustered microcalcifications: automated feature analysis and classification, Radiology, vol. 198, pp. 671–678.
Jiang, Y. Nishikawa, R.M. Schmidt, R.A. Metz, C.E. Giger, M.L. Doi, K. 1998, Benefits
of computer-aided diagnosis (CAD) in mammographic diagnosis of malignant and benign clustered microcalcifications. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography 1998, pp. 215–220. Kluwer, Dordrecht.
Jiang, Y. Nishikawa, R.M. Schmidt, R.A. Metz, C.E. Giger, M.L. and Doi, K. 1999,
Improving breast cancer diagnosis with computer-aided diagnosis, Academic
Radiol., vol. 6, pp. 22–33. Jirari, M. 2005, "A Computer Aided Detection System for Digital Mammograms
Based on Radial Basis Functions and Feature Extraction Techniques", Proceedings of the 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference Shanghai, China, September 1-4, pp. 4457-4460.
Joachims, T. 1998. Text Categorization with Support Vector Machines: Learning with
Many Relevant Features, in Proc. of the 10th European Conference on Machine Learning, Springer Verlag, Heidelberg, pp. 137–142.
Juhl. J. 1982. Paul and Juhl’s Essentials of Roentgen Interpretation. 4th edition,
Harper & Row, Philadelphia, pp. 340–345. Karssemeijer, N. 1992, A stochastic model for automated detection of calcifications
in digital mammograms, Image and Vision Computing, vol. 10, pp. 369–375. Karssemeijer, N. 1993, Recognition of clustered microcalcifications using a random
field model. In Acharya, R.S. Goldgof, D.B. editors, Biomedical Image
Processing and Biomedical Visualization, vol. 1905, pp. 776–786. Karssemeijer, N. and te Brake, G.M. 1996, Detection of stellate distortions in
mammograms, IEEE Transactions on Medical Imaging, vol. 15, pp. 611–619. Karssemeijer. N. and te Brake, G. 1998 Combining single view features and
asymmetry for detection of mass lesions. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography, pp. 95–102, Nijmegen: Kulwer, Academic Publishers,
297
Karssmeijer, N. Otten, J.D. Verbeek, A.L. Groenewoud, J.H. De Koning, H.J. Hendriks, J.H Holland, R. 2003. Computer-aided detection versus independent double reading of masses on mammograms. Radiology, vol. 227, no. 1, pp. 192-200.
Karahaliou, A. Boniatis, I. Skiadopoulos, S. Sakellaropoulos, P. Likaki, E.
Panayiotakis, G. Costaridou, C. 2006. A texture analysis approach for
characterizing microcalcifications on mammograms, in Proc. of the International Special Topic Conference on Information Technology in Biomedicine, Greece.
Karahaliou, A. Boniatis, I. Sakellaropoulos, P. Skiadopoulos, S. Panayiotakis, G.
Costaridou, L. 2007. Can texture of tissue surrounding microcalcifications in mammography be used for breast cancer diagnosis, Nuclear Instruments and
Methods in Physics Research, vol. 580, no. 2, pp. 1071–1074. Karahaliou, A. Boniatis, I. Skiadopoulos, G. Sakellaropoulos, F. Arikidis, N. Likaki,
E.A. Panayiotakis, G. Costaridou, L. 2008. Breast cancer diagnosis: Analyzing texture of tissue surrounding microcalcifications, IEEE Transactions on
Information Technology in Biomedicine, vol. 12, no. 6, pp. 731–738. Kaufman, L. Rousseeuw, P. 1990. Finding Groups in Data: An Introduction to Cluster
Analysis. J. Wiley, New York. Kaufmann, G.H. Salfity, M.F. Granitto, P. Ceccatto, H.A. 2001, Automated detection
and classification of clustered microcalcifications using morphological filtering and statistical techniques. In Yaffe, M.J. editor, Digital Mammography
2000, in Proc. of the 5th International Workshop on Digital Mammography, Madison, Medical Physics Publishing, pp. 253–258.
Kearns, M. Mansour, Y. Ron, D. 1997. An experimental and theoretical comparison
of model selection methods, Machine Learning, vol. 27, pp. 7–50. Kegelmeyer, W.P. Pruneda, J.M. Bourland, P.D. Hillis, A. Riggs, M.W. Nipper, M.L.
1994, Computer-aided mammographic screening for spiculated lesions, Radiology, vol. 191, pp. 331–337.
Kegelmeyer Jr., W.P. 1994. Evaluation of stellate lesion detection in a standard
mammogram data set, Int. J. Pattern Recogn. Artificial Intell., vol. 7, no. 12, pp. 1477–1493.
298
Khuzi, A.M. Besar, R. Zaki, W.M.D.W. Ahmad, N.N. 2009, Identification of masses in digital mammogram using gray level co-occurrence matrices, Biomedical
Imaging and Intervention Journal, vol. 5, no. 3, pp. 1-13. Kim, J.K.K. Min, B. 2002, Classification of Malignant and Benign Tumors Using
Boundary Characteristics in Breast Ultrasonograms, Journal of Digital
Imaging, vol. 15, pp. 224-227. Kobatake, H. Takeo, H. Nawano, S. 1998, Microcalcification detection system for
full-digital mammography. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography 1998, pp. 201–204. Kluwer, Dordrecht.
Kocur, C.M. Rogers, S.K. Myers, L.R. Bums, T. Kabrisky, M. Hoffmeister, J.W. Bauer,
K.W. Steppe, J.M. 1996, Using neural networks to select wavelet features for breast cancer diagnosis, IEEE Engineering In Medicine And Biology Magazine, May/June, pp. 95-105.
Kohavi, R. 1995. A Study of cross-validation and bootstrap for accuracy estimation
and model selection, in Proc. of the International Joint Conference on Artificial Intelligence, pp. 1137–1145.
Kohavi, R. and John, G. 1997. Wrappers for feature subset selection, Artificial
Intelligence Journal, vol. 97, pp. 273–324. Koller, D. and Sahami, M. 1996, Toward Optimal Feature Selection, in Proc. of the
Thirteenth International Conference on Machine Learning, ICML, pp. 284–292.
Kopans, D.B. 1989, Breast Imaging. J. B. Lippincott Company. Kopans D.B. 1992, The positive predictive value of mammography, Amer. J.
Roentgenol., vol. 158, pp. 521–526, 1992. Kramer, D. and Aghdasi, F. 1999, Texture analysis techniques for the classification of
microcalcifications in digitized mammograms, in Proc. of the 5th IEEE AFRICON Conference, Cape Town, Africa, pp. 395–400.
Krupinski, E.A. Nodine, C.F. 1994, Gaze duration predicts the locations of missed
lesions in mammography. In Gale, A.G. Astley, S.M. Dance, D.R. Cairns, A.Y. editors, Digital Mammography, pp. 399–405. Elsevier, Amsterdam.
299
Kundel, H.L. Nodine, C.F. 1978, Studies of eye movements and visual search in radiology. In Seders, J.A.W. Fisher, D. Monty, R. editors, Eye movements and
the higher psychological functions. Hillsdale, New Jersey. Kwok, S. Z. Chandrasekhar, R. Attikiouzel, Y. and Rickard, M.T. 2004, Automatic
pectoral muscle segmentation on mediolateral oblique view mammograms, IEEE Transactions on Medical Imaging, vol. 23, no. 9, pp. 1129–1140.
Lai, S.M. Li, X. Bischof, W.F. 1989, On techniques for detecting circumscribed
masses in mammograms, IEEE Trans. Medical Imaging, vol. 8, no. 4, pp. 377–386.
Laming, D. 1995, Screening cervival smears, British Journal of Psychology, vol. 86,
pp. 507–516. Lanyi, M. 1988, Breast calcifications: Springer-Verlag, Berlin, Heidelberg. Lasdon, L.S. 1970. Optimization Theory for Large Systems, MacMillan.3 Lau, T.K. and Bischof, W.F. 1991, Automated detection of breast tumors using the
asymmetry approach, Comp and Biomed Research, vol. 24, pp. 273–295. Laws, K.I. 1980, Textured Image Segmentation. PhD thesis, University of Southern
California, United States. LeCun, Y. 1986. Learning processes in an asymmetric threshold network. In
Disordered Systems and Biological Organizations, pp. 233–240, Springer-Verlag, Les Houches, France.
Lee, S.K. Lo, C.S. Wang, C.M. Chung, P.C. Chang, C.I. Yang, C.W. Hsu, P.C. 2000, A
computer-aided design mammography screening system for detection and classification of microcalcifications, International Journal of Medical
Informatics, vol. 60, pp. 29–57. Lee, R. Alberdi, E. Taylor, P. 2001, A comparative study of four techniques for
calcification detection. In Yaffe, M.J. editor, Digital Mammography 2000, in Proc. of the 5th International Workshop on Digital Mammography, Madison, Medical Physics Publishing, pp. 264–271.
300
Lerner, B.H. 2002, When statistics provide unsatisfying answers: revisiting the breast selfexamination controversy, Canadian Medical Association Journal, vol. 166, no. 2, pp. 199-201.
Levi, F. Lucchini, F. Negri, E. Boyle, P. La and Vecchia, C. 2003, ‘Mortality from
major cancer sites in the European Union’, Annals of Oncology, vol. 14, pp. 490-495.
Levi, F. Bosetti, C. Lucchini, F. Negri, E. La Vecchia, C. 2005, Monitoring the decrease
in breast cancer mortality in Europe, European Journal of Cancer Prevention, vol. 14, no. 6, pp. 497–502.
Lewin, J.M. Hendrick, R.E. D'orsi, C.J. Isaacs, P.K. Moss, L.J. Karellas, A. Sisney, G.A.
Kuniu, C.C. Cutter, G.R. 2001. Comparison of full-field digital mammography with screen-film mammography for cancer detection: results of 4,945 paired examinations, Radiology, vol. 218, no. 3, pp. 873-880.
Li, H. Lui, K.J. Lo, S.C.B. 1997, Fractal modelling and segmentation for the
enhancement of microcalcifications in digital mammograms, IEEE
Transactions on Medical Imaging, vol. 16, no. 6, pp. 785–798. Liang, N.-Y. Huang, G.-B. Saratchandran, P. Sundararajan, N. 2006, A Fast and
Accurate Online Sequential Learning Algorithm for Feedforward Networks, IEEE Transactions on Neural Networks, vol. 17, no. 6, pp. 1411–1423.
Lim, J. S. 1990. Two-Dimensional Signal and Image Processing, Englewood Cliffs,
NJ, Prentice Hall, pp. 469–476. Lim, G.C.C. Rampal, S. Halimah, Y. (eds.) 2008, Cancer Incidence in Peninsular
Malaysia 2007-2008. National Cancer Registry, Kuala Lumpur. Lloyd, S.P. 1982, Least squares quantization in PCM, IEEE Transactions on
Information Theory, vol. 2, pp. 129–137. Lladó, X. Oliver, A. Martì, J. Freixenet, J. 2007. Dealing with false positive reduction in
mammographic mass detection, in Proc. of the National Conference on Medical Image Understanding and Analysis UK, pp. 81–85, Aberystwyth, Wales, UK.
Lladó, X. Oliver, A. Freixenet, J. Martì, R. Martì, J. 2009. A textural approach for mass
false positive reduction in mammography, Computerized Medical Imaging and
Graphics, vol. 33, no. 6, pp. 415–422.
301
Lu, S. Bottema, M.J. 2001, Classifying lobular and DCIS microcalcification. In Yaffe, M.J. editor, Digital Mammography 2000, in Proc. of the 5th International Workshop on Digital Mammography, Madison, Medical Physics Publishing, pp. 280–284.
Lyra, M. Lyra, S. Kostakis, B. Drosos, S. Georgosopoulos, C. 2008. Digital
mammography texture analysis by computer assisted image processing, in Proc. of the IEEE International Workshop on Imaging, Chania, Greece, pp. 223–227.
Makinacı, M. 2005, Support Vector Machine Approach for Classification of
Cancerous Prostate Regions, World Academy of Science, Engineering and
Technology, no. 7, August, pp. 166-169. Manning, D.J. Ethell, S.C. Donovan, T. 2004. Detection or decision errors? Missed
lung cancer from the posteroanterior chest radiograph. Br. J. Radiol., vol. 77, no. 915, pp. 231-235.
Manzano-Lizcano, J. A. Sánchez-Ávila, C. Moyano-Pérez. L. 2004, "A
Microcalcification Detection System for Digital Mammography using the Contourlet Transform", in Proceedings of the 2004 International Conference on Computational & Experimental Engineering & Science, 26-29 July, Madeira, Portugal, pp. 611-616.
Mao, J. Jain, A.K. 1996. A self-organizing network for hyperellipsoidal clustering
(HEC). IEEE Transactions on Neural Networks, vol. 7, pp. 16–29. Martins, L.de.O. Junior, G.B. Silva, A.C. Paiva, A.C.de. Gatass, M. 2009, Detection of
Masses in Digital Mammograms using K-means and Support Vector Machine, Electronic Letters on Computer Vision and Image Analysis, vol. 8, no. 2, pp. 39-50.
Martì, R. Zwiggelaar, R. Rubin, C. 2000. A novel similarity measure to evaluate image
correspondence, in Proc. of the 15th International Conference on Pattern Recognition, vol. 3, pp. 3171–3174.
Martí, J. Freixenet, J. García, R. Español, J. Golobardes, E. Salamó, M. 2001,
Classification of microcalcifications in digital mammograms using case-based reasoning. In Yaffe, M.J. editor, Digital Mammography 2000, in Proc. of the 5th International Workshop on Digital Mammography, Madison, Medical Physics Publishing, pp. 285–294.
302
Martì, J. Freixenet, J. Noz, X.M. Oliver, A. 2003. Active region segmentation of
mammographic masses based on texture, contour and shape features. Springer-Verlag Berlin Heidelberg, Lecture Notes on Computer Science, vol. 2652, pp. 478–485.
Meesman, D. Scheunders, P. VanDyck, D. 1998, Classification of microcalcifications
using texture-based features. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography 1998, pp. 233–236. Kluwer, Dordrecht.
Mercer, J. 1909. Functions of positive and negative type and their connection with
the theory of integral equations, Transactions of the London Philosophical
Society, vol. 29, pp. 415–446. Miller, P. Astley, S. 1992. Classification of breast tissue by texture analysis, Image
and Vision Computing, vol. 10, pp. 277–283. Miller, A.B. Baines, C.J. To, T. Wall, C. 1992a. Canadian National Breast Screening
Study: 1. Breast cancer detection and death rates among women aged 40 to 49 years, CMAJ, vol. 147, no. 10, pp. 1459-1476.
Mirzaalian, H. Ahmadzadeh, M.R. and Sadri, S. 2007, Pectoral Muscle Segmentation
on Digital Mammograms by Nonlinear Diffusion Filtering, in Proc. of IEEE Pacific Rim Conference on Communications, Computers and Signal Processing, 22-24 Aug., pp. 581–584.
Monsees, B.S. 1995, Evaluation of Breast Microcalcifications, The Radiologic Clinics
of North America, Breast Imaging, vol. 33, no. 6, pp. 1109-1121. Morrow, W.M. Paranjape, R.B. Rangayyan, R.M. Desautels, J.E.L., 1992, "Region-
based contrast enhancement of mammograms", IEEE Transactions on Medical
Imaging, vol. 11, no. 3, pp. 392–406. Moskowitz, M. 1989, Impact of a priory medical detection on screening for breast
cancer, Radiology, vol. 184, pp. 619–622. Mousa, R. Munib, Q. Moussa, A. 2005, Breast cancer diagnosis system based on
wavelet analysis and fuzzy-neural. Expert Systems With Applications, vol. 28, no. 4, pp. 713-723.
303
Mudigonda, N.R. Rangayyan, R. Desautels, J.E.L. 2000, Gradient and Texture Analysis for the Classification of Mammographic Masses, IEEE Transactions
on Medical Imaging, vol. 19, no. 10, pp. 1032-1043. Mudigonda, N.R. Rangayyan, R.M. and Desautels, J.E.L. 2001, Detection of Breast
Masses in Mammograms by Density Slicing and Texture Flow-Field Analysis, IEEE Transactions on Medical Imaging, vol. 20, no. 12, pp. 1215-1227.
Mutihac, R. Colavita, AA. Cicuttin, A. Cerdeira, A. 1998, Maximum entropy
improvement of x-ray digital mammograms. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography, pp. 329–336. Kluwer, Dordrecht.
Müller, K.-R. Mika, S. Ratsch, G. Tsuda, K. Scholkopf, B. 2001, An introduction to
kernel-based learning algorithms, IEEE Trans on Neural Networks, vol. 12, no. 2, pp. 181–201.
Nagel, R.H. Nishikawa, R.M. Papaioannou, J. Doi, K. 1998, Analysis of methods for
reducing false positives in the automated detection of clustered microcalcifications in mammograms, Med. Phys., vol. 25, no. 8, pp. 1502–1506.
Netsch, T. Biel, M. Peitgen, H.O. 1998, Display of high-resolution digital
mammograms on crt monitors. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography, pp. 313–320. Kluwer, Dordrecht.
Newcomb, P.A. Weiss, N.S. Storer, B.E. Scholes, D. Young, B.E. Voigt, L.F. 1991,
Breast self-examination in relation to the occurrence of advanced breast cancer, J Nat Cancer Inst, vol. 83, no. 4, pp. 260–265.
Ng, S.L. and Bischof, W.F. 1992, Automated detection and classification of breast
tumors, Comput Biomed Res., vol. 25 pp. 218–237 Nicolaou, N. Petroudi, S. Georgiou, J. Polycarpou, M. and Brady, M. 2008, Digital
mammography: Towards pectoral muscle removal via Independent Component
Analysis, in Proc. of 4th IET International Conference on Advances in Medical, Signal and Information Processing, 14-16 Jul., pp. 1–4.
304
Nishikawa, R.M. Giger, M.L. Doi, K. Vyborny, C.J. Schmidt, R.A. 1994, Computer-aided detection and diagnosis of masses and clustered microcalcifications from digital mammograms. In Bowyer, K.W. Astley, S.M. editors, State of the
art in digital mammographic image analysis, vol. 9 of series in machine perception and artificial intelligence, World Scientific, pp. 82–102.
Novikoff, A.B.J. 1962. On convergence proofs on perceptrons, in Proc. of the
Symposium on the Mathematical Theory of Automata, vol. 12, pp. 615–662. Oliver, A. Freixenet, J. Bosch, A. Raba, D. Zwiggelaar, R. 2005. Automatic
classification of breast tissue. Springer-Verlag, Berlin, Heidelberg, pp 431–438.
Oliver, A. Lladó, X. Martì, R. Freixenet, J. Zwiggelaar, R. 2007a. Classifying
mammograms using texture information, in Proc. of the National Conference on Medical Image Understanding and Analysis UK, pp. 223–227, Aberystwyth, Wales, UK.
Oliver, A. Lladó, X. Martì, R. Freixenet, J. Martì, J. 2007b. False positive reduction in
mammographic mass detection using local binary patterns, in Proc. of the Int. Conf. Med. Image Comput. Comput. Assist. Interv., vol. 4478, pp. 286–293.
Olsen, O. Gotzsche, P.C. 2001. Cochrane review on screening for breast cancer with
mammography. Lancet, vol. 358, pp. 1340-1342. Osareh, A. Mirmehdi, M. Thomas, B.T. Markham, R. 2002, Comparative Exudate
Classification using Support Vector Machines and Neural Networks, in Proc. of the 5th International Conference on Medical Image Computing and Computer-Assisted Intervention, Tokyo, Japan, pp. 413–420.
Oskoei, M.A. Hu, H. 2008. Support Vector Machine-Based Classification Scheme for
Myoelectric Control Applied to Upper Limb, IEEE Transactions on Biomedical
Engineering, vol. 55, no. 8, pp. 1956–1965. Osuna, E. 1998. Applying SVMs to Face Detection, IEEE Intelligent Systems, vol. 13,
pp. 23–26. Özekes, S. Osman, O. Çamurcu, A.Y. 2005. Mammographic Mass Detection Using a
Mass Template, Korean Journal of Radiology, vol. 6, no. 4, pp. 221–228.
305
Palmer, G.M. Zhu, C. Breslin, T.M. Xu, F. Gilchrist, K.W. and Ramanujam, N. 2003, Comparison of Multiexcitation Fluorescence and Diffuse Reflectance Spectroscopy for the Diagnosis of Breast Cancer, IEEE Transactions on
Biomedical Engineering, vol. 50, no. 11, pp. 1233-1242. Papadopoulosa, A. Fotiadisb, D.I. Likasb, A. 2005, Characterization of clustered
microcalcifications in digitized mammograms using neural networks and support vector machines, Artificial Intelligence in Medicine, vol. 32, no. 2, pp. 141-150.
Parkin, D. M. Bray, F. Ferlay, J. Pisani, P. 2005, Global Cancer Statistics, 2002, CA: A
Cancer Journal for Clinicians, vol. 55, no. 74. Parr, T.C. Astley, S.M. Taylor, C.J.Boggis, C.R.M. 1996a, Model based classification of
linear structures in digital mammograms. Parr, T.C. Taylor, C.J. Astley, S.M. Boggis, C.R.M. 1996b, A statistical representation
of pattern structure for digital mammography. In Doi, K. Giger, M.L. R M Nishikawa, R.M. Schmidt, R.A. editors, Digital Mammography, pp. 357–360. Elsevier, Amsterdam.
Pavlidis, T. Liow, Y.-T. 1990, Integrating region growing and edge detection, IEEE
Trans. Pattern Anal. Machine Intell., vol. 12, pp. 225–233. Peer, P.G.M. Werre, J.M. Mravunac, M. Hendriks, J.H.C.L. Holland, R. Verbeek, A.L.M.
1995, Effect on breast cancer mortality of biennial mammographic screening of women under age 50, Int. J. Cancer, vol. 60, pp. 808–811.
Pfisterer, R. Aghdasi, R. 1998. Detection of masses in digitised mammograms, in
Proc. of the 1998 South African Symposium on Communications and Signal Processing, South Africa, pp. 115–120.
Pfisterer, R. Aghdasi, F. 1999, Comparison of texture based algorithms for the
detection of masses in digitized mammograms, in Proc. of the IEEE AFRICON, Cape Town, South Africa, vol. 1, pp. 383–388.
Pfisterer, R. Aghdasi, F. 2001. Tumor detection in digitized mammograms by image
texture analysis. Society of Photo-Optical Instrumentation Engineers, vol. 40, pp. 209–216.
306
Pisano, E.G. Gatsonis, C. Hendrick, E. Yaffe, M. Baum, J.K. Acharyya, S. Conant, E.F. Fajardo, L.L. Basett, L. D'orsi, C.J. Jong, R. Rebner, M. 2005. Diagnostic performance of digital versus film mammography for breast-cancer screening, N Engl J Med, vol. 353, no. 17, pp. 1773-1783.
Platt, J.C. 1998. Sequential Minimal Optimization: A Fast Algorithm for Training
Support Vector Machines, Technical Report MSR-TR-98-14, Microsoft Research Center, Redmond, United States.
Platt, J.C. 1999a. Fast Training of Support Vector Machines using Sequential
Minimal Optimization. In Smola, A.J. Bartlett, P.L. Schölkopf, B. Schuurmans, D. editors, Advances in Large Margin Classifiers, pp. 185–208, MIT Press, Cambridge, Massacheusets.
Platt, J.C. 1999b. Probabilistic Output for Support Vector Machines and Comparison
to Regularized Likelihood Methods. In Smola, A.J. Bartlett, P.L. Schölkopf, B. Schuurmans, D. editors, Advances in Large Margin Classifiers, pp. 61–74, MIT Press, Cambridge, Massacheusets.
Pohlman, S. Powell, K.A. Obuchowski, N.A. Chilcote, W.A. Grundfest-Broniatowski,
S. 1996, Quantitative classification of breast tumors in digitized mammograms, Med Phys., vol. 23, pp. 1337–1345.
Polakowski, W.E. Cournoyer, D.A. Rogers, S.K. DeSimio, M.P. Ruck, D.W.
Hoffmeister, J.W. Raines, R.A. 1997, Computer-aided breast cancer detection and diagnosis of masses using difference of gaussians and derivative-based feature saliency, IEEE Trans on Medical Imaging, vol. 16, no. 6, pp. 811–819.
Popli, M.B. 2001, Pictorial essay: Mammographic features of breast cancer, Ind J.
Radiol. Imag., vol. 11, pp. 175–179. Qian, W.C.L. Kallergi. M. Clark, R.A. 1994, Tree-structured nonlinear filters in digital
mammography, IEEE Trans. Medical Imaging, vol. 13, no.1, pp. 25–36. Raba, D. Oliver, A. Martí, J. Peracaula1, M. and Espunya, J. 2005, Breast
Segmentation with Pectoral Muscle Suppression on Digital Mammograms, in Lecture Notes on Computer Science, vol. 3523, pp. 471–478.
Rahbar, G. Sie, A.C. Hansen, G.C. Prince, J.S. Melany, M.L. Reynolds, H.E. Jackson, V.P.
Sayre, J.W. Bassett, L.W. 1999. Benign versus malignant solid breast masses: US differentiation, Radiology, vol. 213, no. 3, pp. 889-894.
307
Rangayyan, R.M. El-Faramawy, N.M. Desautels, J.E.L. Alim, O.A. 1997, Measures of Acutance and Shape for Classification of Breast Tumors, IEEE Transactions on
Medical Imaging, vol. 16, no. 6, pp. 799-810. Rangayyan, R.M. Desautels, J.E.L. 2000, Boundary modelling and shape analysis
methods for classification of mammographic masses, Medical and Biological
Engineering and Computing, vol. 38, no. 5, pp. 487-496. Rangayyan, R.M. 2005. Biomedical Image Analysis. Biomedical Engineering Series.
CRC Press LLC, ISBN 0-8493-9695-6. Richards, W. Polit, A. 1974. Texture matching. Kybernetic, vol. 16, pp. 155–162. Ripley, B.D. 1996, Pattern Recognition and Neural Networks. Cambridge University
Press. Roebuck, E.J. and Blamey, R.W. 1990, Clinical Radiology of the Breast. London:
Heinemann Medical Books. Rosenblatt, F. 1962. Principles of Neurodynamics. Spartan Books, Washintgon DC. Rumelhart, D.E. Hinton, G.E. Williams, R.J. 1986. Learning internal representations
by error propagaton. Parallel Distributed Processing: Explorations in the
Microstructure of Cognition, vol. 1, pp. 318–362. Rumelhart, D.E. Hinton, G.E. Williams, R.J. 1986, Learning representations by back-
propagating errors, Nature, vol. 323, pp. 533–536. Sahiner, B. Chan, H.P. Petrick, N. Wei, D. Helvie, M.A. Adler, D.D. Goodsitt, M.M.
1996, Classification of mass and normal breast tissue: a convolution neural network classifier with spatial domain and texture images, IEEE Trans Med
Imag., vol. 15, pp. 598–610. Sallam, M. and Bowyer, K.W. 1994, Registering time-sequences of mammograms
using a two-dimensional unwarping technique. In Gale, A.G. Astley, S.M. Dance, D.R. Cairns, A.Y. editors, Digital Mammography, pp. 121–131. Elsevier, Amsterdam.
Sample, J.T. 2005, Computer Assisted Screening Of Digital Mammogram Images. PhD
thesis, Dept. of Computer Science, Louisiana State University, USA.
308
Savage, C.J. Gale, A.G. Pawley, E.F. Wilson, A.R.M. 1994, To err is human; to compute divine? In Gale, A.G. Astley, S.M. Dance, D.R. Cairns, A.Y. editors, Digital
Mammography, pp. 405–414. Elsevier, Amsterdam. Schmidt, F. Hartwagner, K.A. Spork, E.B. Groell, R. 1998, Medical audit after 26, 711
breast imaging studies: improved rate of detection of small breast carcinomas, Cancer, vol 83, no. 12, pp. 2516–2520.
Schmidt, R.A. Nishikawa, R.M. Osnis, R.B. Schreibman, K.L. Giger, M.L. Doi, K. 1996,
Computerized detection of lesions missed by mammography. In Doi, K. Giger, M.L. Nishikawa, R.M. Schmidt, R.A. editors, Digital Mammography, pp. 105–110. Elsevier, Amsterdam.
Schölkopf, B. 1997. Support Vector Learning. PhD thesis, Technischen Universität
Berlin. Published by: R. Oldenbourg Verlag, Munich. Selvan, S.E. Xavier, C.C. Karssemeijer, N. Sequiera, J. Cherian, R.A. Dhala, B.Y. 2006.
Parameter Estimation in Stochastic Mammogram Model by Heuristic Optimization Techniques, IEEE Trans. on Information Technology in
Biomedicine, vol. 10, no. 4, pp. 685–695. Shapiro, S. Strax, P. Venet, L. 1971, Periodic breast cancer screening in reducing
mortality from breast cancer, JAMA, vol. 215, no. 11, pp. 1777–1785. Shapiro, S. Venet, W. Strax, P. Venet, L. Roeser, R. 1982, Ten- to fourteen-year effect
of screening on breast cancer mortality, Journal of the National Cancer
Institute, vol. 69, pp. 349-55. Sheshadri, H.S. Kandaswamya, A. 2007. Experimental investigation on breast tissue
classification based on statistical feature extraction of mammograms, Computerized Medical Imaging and Graphics, vol. 31, no. 1, pp. 46–48.
Sickles, E.A. 1984, Mammographic features of early breast cancer, American Journal
of Roentgenology, vol. 143, pp. 461-464. Sickles, E.A. 1986, Mammographic features of 300 consecutive nonpalpable breast
cancers, American Journal of Roentgenology, vol. 146, pp. 661-663. Sickes, E.A. 1997, Breast cancer screening outcomes in women ages 40-49: clinical
experience with service screening using modern mammography, Journal of
the National Cancer Institute: Monographs, vol. 22, pp. 99-104.
309
Smith, R.A. Caleffi, M. Albert, U.S. Chen, T.H.H. Duffy, S.W. Franceschi, D. Nystrom, L. 2006, Breast Cancer in Limited-Resource Countries: Early Detection and Access to Care, Breast Journal vol. 12, pp. 16-26.
Smola, A.J. Schölkopf, B. 2004. A Tutorial on Support Vector Regression, Statistics
and Computing, vol. 14, no. 3, 2004, pp. 199–222. Soh, L.-K. Tsatsoulis, C. 1999. Texture Analysis of SAR Sea Ice Imagery Using Gray
Level Co-Occurrence Matrices, IEEE Transactions on Geoscience and Remote
Sensing, vol. 37, no. 2, pp. 780–795. Soltanian-Zadeh, H. Rafiee-Rad, F. and Pourabdollah-Nejad, S. 2004, ‘Comparison of
multiwavelet, wavelet, Haralick, and shape features for microcalcification classification in mammograms’, Pattern Recognition, vol. 37, no. 10, pp. 1973–1986.
Song, E. Jiang, L. Jin, R. Zhang, L. Yuan, Y. Li, Q. 2009. Breast Mass Segmentation in
Mammography Using Plane Fitting and Dynamic Programming, Academic
Radiology, vol. 16, no. 7, pp. 826–835. Sorantin, E. Schmidt, F. Mayer, H. Winkler, P. Szepesvari, C. Graif, E. Schuetz, E.
1998, Automated detection and classification of microcalcifications in mammograms using artificial neural nets. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital Mammography, pp. 225–232. Kluwer, Dordrecht.
Statistics Toolbox (R2009b), MATLAB (Matrix Laboratory). Mathworks, 2009.
[Online]. Available at: http://www.mathworks.com/help/toolbox/stats/ Strickland, R.N. 1996a, Wavelet transform for detecting microcalcifications in
mammograms, IEEE Transactions on Medical Imaging, vol. 15, no. 2, pp. 218–229.
Strickland, R.N. Baig, L.J. Dallas, W.J. Krupinski, E.A. 1996, Wavelet-based image
enhancement as an instrument for viewing cad data. In Doi, K. Giger, M.L. Nishikawa, R.M. Schmidt, R.A. editors, Digital Mammography, pp. 441–446. Elsevier, Amsterdam.
Stoutjesdijk, M.J. Boetes, C. Jager, G.J. Beex, L. Bult, P. Hendriks, J.H. Laheij, R.J.
Massuger, L. Van Die, L.E. Wobbes, T. Barentsz, J.O. 2001. Magnetic resonance imaging and mammography in women with a hereditary risk of breast cancer, J Natl Cancer Inst, vol. 93, no. 14, pp. 1095–1102.
310
Subashini, T.S. Ramalingama, V. Palanivela, S. 2010. Automated assessment of breast tissue density in digital mammograms, Computer Vision and Image
Understanding, vol. 114, no. 1, pp. 33–43. Suckling, J. Parker, J. Dance, D. et al., 1994. The mammographic image analysis
society digital mammogram database. Exerpta Medica, International Congress
Series, vol. 1069, pp. 375–378. [Online] Available at: http://peipa.essex.ac.uk/info/mias.html
Suzuki, K. Li, F. Sone, S. Doi, K. 2005, Computer-aided diagnostic scheme for
distinction between benign and malignant nodules in thoracic low-dose CT by use of massive training artificial neural network, IEEE Trans on Medical
Imaging, vol. 24, no. 9, pp. 1138–1150. Svetnik, V. Liaw, A. Tong, C. Wang, T. 2004. Application of Breiman’s random forest
to modeling structure-activity relationships of pharmaceutical molecules. In Roli, F. Kittler, J. Windeatt, T., editors, Proceedings of the 5th International Workshopon Multiple Classifier Systems, Lecture Notes in Computer Science vol. 3077, pp. 334–343, Springer.
Tabár, L. Fagerbert C.J.G, Gad, A. Baldetorp, L. Holmberg, L.H. Grontoft, O.
Ljungquist, U. Lundstrom, B. Manson, J.C. Erklung, G. Day, N.E. and Pettersson F. 1985, Reduction in mortality from breast cancer after mass screening with mammography. Randomized trial from the breast cancer screening work group of the swedish national board of health and welfare, Lancet, vol. 1, pp. 829–832.
Tabár, L. Fagerberg, G. Day, N.E. Holmberg, L 1987, What is the optimum interval
between mammographic screening examinations? an analysis based on the latest results of the swedish two-county breast cancer screening trial, Br J
Cancer, vol. 55, no. 5, pp. 547–551. Tabár, L. Fagerbert C.J.G, Chen, H.H. Duffy, S.W. Smart, C.R. Gad, A. Smith, R.A. 1995.
Efficacy of breast cancer screening by age. New results from the Swedish Two-County Trial. Cancer vol. 75, no. 10, pp. 2507–2517.
Tarassenko, L. 1998, Guide to Neural Computing Applications, Butterworth-
Heinemann. Thangavel, K. Karnan, M. Kumar, R.S. Mohideen, A.K. 2005a, Automatic Detection of
Microcalcification in Mammograms-A Review. International Journal on
Graphics Vision and Image Processing, vol. 5, no. 5 pp. 31–61.
311
Thangavel, K. and Karnan, M. 2005b, CAD System for Preprocessing and Enhancement Of Digital Mammograms, ICGST International Journal on
Graphics, Vision and Image Processing (GVIP), vol. 5, no. 9, pp. 69–74. Thomas, D.B. Gao, D.L. Ray, R.M. Wang, W.W. Allison, C.J. Chen, F.L. Porter, P. Hu,
Y.W. Zhao, G.L. Pan, L.D. 2002, Randomized Trial of Breast Self-Examination in Shanghai: Final Results, Journal of the National Cancer Institute, vol. 94, no. 19, pp. 1445–1457.
Thurfjell, E.L. Lernevall, K.A. and Taube A.A.S 1994, Benefit of independent double
reading in a population-based mammography screening program, Radiology, vol. 191, pp. 241–244.
Thurfjell, E.L. Lindgren, J.A.L. 1996, Breast cancer survival rates with
mammographic screening: similar favorable survival rates for women younger and those older than 50 years, Radiology, vol. 201, no. 2, pp. 421–426.
Timp, S. Karssemeijer, N. 2004. A new 2D segmentation method based on dynamic
programming applied to computer aided detection in mammography, Med.
Phys., vol. 31, no. 5, pp. 958–971. Tyczynski, J.E. Plesko, I. Aareleid, T. Primic-Zakelj, M. Dalmas, M. Kurtinaitis, J.
Stengrevics, A. Parkin, D.M. 2004, Breast cancer mortality patterns and time trends in 10 new EU member states: Mortality declining in young women, but still increasing in the elderly, International Journal of Cancer, vol. 112, no. 6, pp. 1056–1064.
Underwood, J.C.E. (ed.) 1992. General and systematic pathology, chapter 16.
Churchill Livingstone. University of California Irvine (UCI) Machine Learning Repository, Center for
Machine Learning and Intelligent Systems, University of California (2010). [Online]. Datasets available at: http://archive.ics.uci.edu/ml/
Vaino, H. Bianchini, F. (eds.) 2002. IACR handbooks of cancer prevention. Lyon:
IARCPress. Van Dijck, J.A.M. Verbeek, L.M. Hendriks, J.H.C.L. Holland, R. 1993, The current
detectability of breast cancer in a mammographic screening program, Cancer, vol. 72, pp. 1933–1938.
312
Van Dijck, J.A. Verbeek, A.L. Beex, L.V. Hendriks, J.H.C.L. Holland, R. Mravunac, M. Straatman, H. Werre, J.M. 1997, Breast-cancer mortality in a non-randomized trial on mammographic screening in women over age 65, Int J Cancer, vol. 70, no. 2, pp. 164–168.
Vapnik, V. Chervonenkis, A.J. 1968. On the uniform convergence of relative
frequencies of events to their probabilities, Dokaldy Akademii Nauk USSR. Vapnik, V. Chervonenkis, A.J. 1974. Theory of Pattern Recognition (in Russian).
Nauka, Moscow, (German Translation: W. Wapnik & A. Tscherwonenkis, Theorie der Zeichenerkennung, Akademie–Verlag, Berlin, 1979.
Vapnik, V. 1979. Estimation of Dependences Based on Empirical Data (in Russian).
Nauka, Moscow, (English translation: Springer Verlag, New York, 1982). Vapnik, V. 1995. The Nature of Statistical Learning Theory: Springer Verlag. Vapnik, V. 1998. Statistical Learning Theory: John Wiley. Varela, C. Karssemeijer, N. Tahoces, P. 2001. Classification of breast tumors on
digital mammograms using laws texture features. Springer-Verlag, Berlin, Heidelberg, Lecture Notes on Computer Science, vol. 2208, pp. 1391–1392.
Varela, C. Timp, S. Karssemeijer, N. 2006, Use of border information in the
classification of mammographic masses, Phys. Med. Biol., vol. 51, no.2, pp. 425-441.
Veldkamp, W.J.H. Karssemeijer, N. 1998, Improved correction for signal dependent
noise applied to automatic detection of microcalcifications. In Karssemeijer, N. Thijssen, M.A.O. Hendriks, J.H.C.L. van Erning, L.J.T.O. editors, Digital
Mammography 1998, pp. 169–176. Kluwer, Dordrecht. Veldkamp, W.J.H. Karssemeijer, N. Otten, J.D.M. Hendriks, J.H.C.L. 2000, Automated
Classification of Clustered Microcalcifications into Malignant and Benign Types, Medical Physics, vol. 27, no. 11, pp. 2600–2608.
Verma, B. and Zakos, J. 2001, A Computer-Aided Diagnosis System for Digital
Mammograms Based on Fuzzy-Neural and Feature Extraction Techniques, IEEE Transactions on Information Technology in Biomedicine, vol. 5, no. 1, pp. 46-54.
313
Veropoulos, K. 2001, Machine Learning Approaches to Medical Decision Making.
PhD thesis, University of Bristol, United Kingdom. Vitak, B. 1998, Invasive interval cancers in the Ӧstergӧtland mammographic
screening programme: Radiological analysis, European Radiology, vol. 8, pp. 639–646.
Wallet, B.C. Solka, J.L. Priebe, C.E. 1997, A method for detecting microcalcifications
in digital mammograms, Journal of Digital Imaging, vol. 10, pp. 136–139. Wang T.C. and Karayiannis, N.B. 1998, Detection of Microcalcifications in Digital
Mammograms Using Wavelets, IEEE Transactions on Medical Imaging, vol. 17, no. 4, pp. 500–509.
Wang, H.-Q. Huang, D.-S. Wang, B. 2005. Optimisation of Radial Basis Function
Classifiers using Simulated Annealing Algorithm for Cancer Classification, Electronics Letters, vol. 41, no. 11, pp. 630–632.
Wei, D. Chan, H.P. Helvie, M.A. Sahiner, B. Petrick, N. Adler, D.D. Goodsit, M.M. 1995,
Classification of mass and normal breast tissue on digital mammograms: multiresolution texture analysis, Med Phys, vol. 22, pp. 1501–1513.
Wei, L. Yang, Y. Nishikawa, R.M. Jiang, Y. 2005, A study on several machine-learning
methods for classification of malignant and benign clustered microcalcifications, IEEE Transactions on Medical Imaging, vol. 24, no.5 pp. 371-380.
Wen, J. Zhao, J.L. Luo, S.W. Han, Z. 2000, The Improvements of BP Neural Network
Learning Algorithm, in Proc. of the 5th International Conference on Signal Processing Proceedings, Beijing, China, vol. 3, pp. 1647–1649.
Wirth, M.A. Lyon, J. Nikitenko, D. Stapinski, A. 2004, Removing Radiopaque
Artifacts from Mammograms using Area Morphology, in Proc. of the of SPIE Medical Imaging: Image Processing, pp. 1054–1065.
Wirth, M. Nikitenko, D. Lyon, J. 2007, Segmentation of the Breast Region in
Mammograms using a Rule-Based Fuzzy Reasoning Algorithm, ICGST
International Journal on Graphics, Vision and Image Processing (GVIP), vol. 9, no. 5, pp. 13–22.
314
Wolberg, W.H. Mangasarian, O.L. 1990, Multisurface Method of Pattern Separation
for Medical Diagnosis Applied to Breast Cytology, in Proc. of the National Academy of Sciences, U.S.A., vol. 87, Dec. 1990, pp. 9193–9196.
Woods, K. Bowyer, K. 1996, A general view of detection algorithms. In Doi, K. Giger,
M.L. Nishikawa, R.M. Schmidt, R.A. editors, Digital Mammography, pp. 385–390, Elsevier, Amsterdam.
Wu, Y. Giger, M.L. Doi, K. Vyborny, C.J. Schmidt, R.A. and Metz, C.E. 1993,
Application of neural networks in mammography: applications in decision making in the diagnosis of breast cancer, Radiology, vol. 187, pp. 81–87.
Wu, T.F. Lin, C.-J. Weng, R.C. 2004. Probability Estimates for Multiclass
Classification by Pairwise Coupling, Journal of Machine Learning Research, vol. 5, pp. 975–1005.
Xu, W. Li, L. and Liu, W. 2007, A Novel Pectoral Muscle Segmentation Algorithm
Based on Polyline Fitting and Elastic Thread Approaching, in Proc. of the 1st International Conference on Bioinformatics and Biomedical Engineering, 6-8 Jul., pp. 837–840.
Yapa, R.D. Harada, K. 2008, Breast Skin-Line Estimation and Breast Segmentation in
Mammograms using Fast-Marching Method, International Journal of
Biomedical Sciences, vol. 3, no. 1, pp 54–62. Yin, F.F. Giger, M.L. Doi, K. Metz, C.E. Vyborny, C.J. Schmidt, R.A. 1991,
Computerized detection of masses in digital mammograms: Analysis of bilateral substraction images, Med Phys, vol. 18 pp. 955–963.
Yin, F.F. Giger, M.L. Vyborny, C.J. Doi, K. Schmidt, R.A. 1993, Comparison of
bilateralsubstraction and single-image processing techniques in the computerized detection of mammographic masses. Invest Radiol, vol. 6, pp. 473–481.
Yoshida, H. Doi, K. Nishikawa, R.M. 1994, Automated detection of clustered
microcalcifications in digital mammograms using wavelet transform techniques, in Proc. of SPIE 2167: Medical Imaging, pp. 868–886.
Yoshida, H. Doi, K. Nishikawa, R. Giger, M. and Schmidt, R. 1996, An improved
computer-assisted diagnostic scheme using wavelet transform for detecting clustered microcalcifications in digital mammograms, Acad. Radioogy, vol. 3, no. 8, pp. 621–627.
315
Youssry, N. Abou-Chadi, F.E.Z. El-Sayad, A.M. 2003. Early detection of masses in
digitized mammograms using texture features and neuro-fuzzy model, in Proc. of the 4th Annual IEEE Conf on Information Technology Applications in Biomedicine, Birmingham, United Kingdom, pp. 226–233.
Zhang, M. and Giger, M.L. 1995, Automated detection of spiculated lesions and
architectural distortions in digitized mammograms, SPIE 2434, pp. 846–855. Zhao, D. Shridhar, M. Daut, D.G. 1992, Morphology on detection of calcifications in
mammograms, in Proc. the IEEE International Conference on Acoustics, Speech, and Signal Processing, San Francisco, vol. 3, pp. 129–132.
Zheng, B. Chang, Y.H. Gur, D. 1995, Computerized detection of masses in digitized
mammograms using single image segmentation and a multi-layer topographic feature analysis, Acad Radiol., vol. 2, pp. 959–966.
Zonderland, H.M. Coerkamp, E.G. Hermans, J. Van De Vijver, M.J. Van Voorthuisen,
A.E. 1999. Diagnosis of breast cancer: contribution of US as an adjunct to mammography, Radiology, vol. 213, no. 2, pp. 413-422.
Zucker, S.W. Kant, K. 1981. Multiple-level representations for texture discrimination,
in Proc of the IEEE Conference on Pattern Recognition and Image Processing, pp. 609–614.
Zwiggelaar, R. Parr, T.C. Schumm, J.E. Hutt, I.W. Taylor, C.J. Astley, S.M. Boggis,
C.R.M. 1999, Model-based detection of spiculated lesions in mammography. Medical Image Analysis, vol. 3, no. 1, pp. 39–62.
316
APPENDICES
317
APPENDIX A
Data Modeling and Analysis
Figure A.1: GLCM texture features calculated from the malignant ROIs
318
Figure A.2: F-scores of the 1152 GLCM texture feature values
319
Figure A.3: Optimal subset of features obtained using the “F-score + RF + SVM”
technique
320
Figure A.4: Texture features values for optimum subset of 1056 features
321
Figure A.5: Normalized feature values in the range between 0 and 1
322
Figure A.6: Training data feature file with malignant and benign class labels
323
APPENDIX B
SVM Training and Testing
Figure B.1: The LIBSVM SVM training function in MATLAB
324
Figure B.2: SVM model parameters in MATLAB after SVM training
325
Figure B.3: SVM model generated after SVM training in MATLAB
326
Figure B.4: The LIBSVM prediction function in MATLAB
327
Figure B.5: SVM testing and validation in MATLAB using LIBSVM
328
APPENDIX C
LIBSVM Copyright Notice
The breast cancer detection system developed in this research is collaborated with
the Faculty of Computer Science and Information Technology (FSKTM), University
of Malaya and the Department of Radiology, University of Malaya Medical Centre
(UMMC), Kuala Lumpur. The system developed incorporates a tool “LIBSVM”, a
library for support vector machines, which was developed by Chih-Chung Chang
and Chih-Jen Lin. Acknowledgement of the LIBSVM copyright is shown as below.
Copyright (c) 2000-2008 Chih-Chung Chang and Chih-Jen Lin All
rights reserved.
Redistribution and use in source and binary forms, with or
without modification, are permitted provided that the
following conditions are met:
1. Redistributions of source code must retain the above
copyright notice, this list of conditions and the following
disclaimer.
2. Redistributions in binary form must reproduce the above
copyright notice, this list of conditions and the following
disclaimer in the documentation and/or other materials
provided with the distribution.
3. Neither name of copyright holders nor the names of its
contributors may be used to endorse or promote products
derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND
CONTRIBUTORS ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES,
INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
329
APPENDIX D
List of Publications
Journal Publications
[1] Jawad Nagi, Keem Siah Yap, Sieh Kiong Tiong, Syed Khaleel Ahmed, Farrukh Nagi,
"Improving SVM-based Nontechnical Loss Detection in Power Utility Using Fuzzy
Inference System", accepted for publication in IEEE Transactions on Power Delivery on 22nd June 2010. Manuscript ID: PESL-00108-2009.R2.
[2] Mohammad Mehdi Badjian, Jawad Nagi, Sieh Kiong Tiong, Keem Siah Yap, Siaw Paw Koh, Farrukh Nagi, “Comparison of Supervised Learning Techniques for Non-
Technical Loss Detection in Power Utility”, submitted to Malaysian Journal of
Computer Science (MJCS) for first review on 9 April 2010.
[3] Farrukh Nagi, Syed Khaleel Ahmed, Jawad Nagi, "Fuzzy Time-Optimal Controller
(FTOC) for Second Order Nonlinear Systems", submitted to IEEE Transactions on
Systems, Man, and Cybernetics: Part B for first review on 30 November 2009. Paper No: SMCB-E-2009-11-1046.
[4] Jawad Nagi, Keem Siah Yap, Sieh Kiong Tiong, Syed Khaleel Ahmed, Farrukh Nagi, "A
Computational Intelligence Scheme for Prediction of the Daily Peak Load", submitted for second review to Applied Soft Computing (ASOC) on 10 August 2010. Manuscript Reference No: ASOC-D-09-00556.
[5] Jawad Nagi, Keem Siah Yap, Sieh Kiong Tiong, Syed Khaleel Ahmed, and Malik Mohammad, “Nontechnical Loss Detection for Metered Customers in Power
Utility Using Support Vector Machines”, IEEE Transactions on Power Delivery, vol. 25, no. 2, pp. 1162–1171, Apr. 2010.
[6] Farrukh Nagi, Logah Perumal, and Jawad Nagi, “A New Integrated Fuzzy Bang-
Bang Relay Control System”, Mechatronics, vol. 19, no. 5, pp. 748–760, Aug. 2009.
Conference Publications
[1] Jawad Nagi, Keem Siah Yap, Farrukh Nagi, Sieh Kiong Tiong, Siaw Paw Koh, Syed
Khaleel Ahmed, “NTL Detection of Electricity Theft and Abnormalities for Large
Power Consumers in TNB Malaysia”, in Proc. of the 2010 IEEE Student Conference on Research and Development (SCOReD) 2010, 14 Dec. 2010, Malaysia, pp. 1–5.
[2] Jawad Nagi, Sameem Abdul Kareem, Farrukh Nagi, and Syed Khaleel Ahmed, “Automated Breast Profile Segmentation for ROI Detection Using Digital
Mammograms”, in Proc. of the IEEE Conference on Biomedical Engineering and
Sciences (IECBES) 2010, 30 Nov. 2010, Kuala Lumpur, Malaysia, pp. 1–6.
330
[3] Jawad Nagi, Tiong Sieh Kiong, Syed Khaleel Ahmed, and Farrukh Nagi, "Prediction
of PVT Properties in Crude Oil Systems Using Support Vector Machines", in Proc. of the 3rd International Conference on Energy and Environment (ICEE) 2009, Dec.
7-8, 2009, Malacca, Malaysia, pp. 1–5.
[4] Jawad Nagi, Keem Siah Yap, Sieh Kiong Tiong, Abdul Malik Mohammad, and Syed
Khaleel Ahmed, “Non-Technical Loss Analysis for Detection of Electricity Theft
using Support Vector Machines”, in Proc. of the 2nd IEEE International Power and Energy Conference (PECon) 2008, Dec. 1-3, 2008, Johor Bahru, Malaysia, pp. 907–912.
[5] Jawad Nagi, Keem Siah Yap, Sieh Kiong Tiong, and Syed Khaleel Ahmed, “Detection
of Abnormalities and Electricity Theft using Genetic Support Vector Machines”,
in Proc. of the IEEE Region 10 Conference (TENCON) 2008, Nov. 19, 2008, Hyderabad, India, pp. 1–6.
[6] Jawad Nagi, Syed Khaleel Ahmed, and Farrukh Nagi, “Pose Invariant Face
Recognition using Hybrid DWT-DCT Frequency Features with Support Vector
Machines”, in Proc. of the 4th International Conference on Information Technology and Multimedia at UNITEN (ICIMu) 2008, Nov. 18-19, 2008, Bandar Baru Bangi,
Selangor, Malaysia, pp. 99–104.
[7] Jawad Nagi, Keem Siah Yap, Sieh Kiong Tiong, and Abdul Malik Mohammad,
“Intelligent System for Detection of Abnormalities and Theft of Electricity using
Genetic Algorithm and Support Vector Machines”, in Proc. of the 4th International Conference on Information Technology and Multimedia at UNITEN (ICIMu) 2008, Nov. 18-19, 2008, Bandar Baru Bangi, Selangor, Malaysia, pp. 122–127.
[8] Jawad Nagi, Keem Siah Yap, Sieh Kiong Tiong, Syed Khaleel Ahmed, and Farrukh
Nagi, “Intelligent Detection of DTMF Tones using a Hybrid Signal Processing
Technique with Support Vector Machines”, in Proc. of the International Symposium on Information Technology (ITSIM) 2008, Aug. 26-28, 2008, Kuala Lumpur, Malaysia, vol. 4, pp. 1–8.
[9] Jawad Nagi, Sieh Kiong Tiong, Yap Keem Siah, and Syed Khaleel Ahmed, “Dual-tone
Multi-frequency Signal Detection using Support Vector Machines”, in Proc. of the 6th National Conference on Telecommunication Technologies and Malaysia
Conference on Photonics (NCTT-MCP) 2008, Aug. 26-28, 2008, Putrajaya, Malaysia, pp. 350–355.
[10] Jawad Nagi, Syed Khaleel Ahmed, and Farrukh Nagi, “Palm Biodiesel an
Alternative Green Renewable Energy for the Energy Demands of the Future”, in Proc. of the International Conference on Renewable Energy and Sustainability (ICCBT) 2008, Jun. 16-20, 2008, Kuala Lumpur, Malaysia, pp. 79–94.
331
[11] Jawad Nagi, Keem Siah Yap, Sieh Kiong Tiong, and Syed Khaleel Ahmed, “Electrical
Power Load Forecasting using Hybrid Self-Organizing Maps and Support Vector
Machines”, in Proc. of the 2nd International Power Engineering and Optimization Conference (PEOCO) 2008, Jun. 4-5, 2008, Shah Alam, Malaysia, pp. 51–56.
[12] Jawad Nagi, Syed Khaleel Ahmed, and Farrukh Nagi, “A MATLAB based Face
Recognition System using Image Processing and Neural Networks”, in Proc. of the 4th International Colloquium on Signal Processing and its Applications (CSPA)
2008, Mar. 7-9, 2008, Kuala Lumpur, Malaysia, pp. 83–88.
332
BIODATA OF THE AUTHOR
Jawad Nagi, was born in Karachi, Pakistan on March
23, 1985. He received his Bachelor’s degree from Universiti Tenaga Nasional (UNITEN), Malaysia with Honors in Electrical and Electronics Engineering in
2007. In 2009 he was awarded the Master of Electrical Engineering degree from UNITEN in 2009.
He is currently pursuing a Master’s of Computer Science degree at University of Malaya, Malaysia,
which is expected to complete in August 2010. He is currently working as a Research Engineer at UNITEN R&D Sdn. Bhd. of Universiti Tenaga Nasional
(UNITEN) since January 2008. He also is involved in teaching activities at the Asia Pacific Institute of
Information Technology (APIIT), Kuala Lumpur, Malaysia. His research interests include pattern recognition, machine learning, image processing, load forecasting, fuzzy logic, neural networks, support vector machines,
robotics and control systems. His publications and resume can be found at: http://metalab.uniten.edu.my/~jawad/papers/
